Multi-nest remote take-off and landing unmanned aerial vehicle autonomous inspection method and system

Abstract

The embodiment of the invention discloses a multi-nest remote take-off and landing unmanned aerial vehicle autonomous inspection method and system, and belongs to the technical field of unmanned aerial vehicles. According to the embodiment of the invention, the target inspection space area is identified, and the related vector and information are acquired, so that the inspection task planning can accurately meet the actual demand. A multi-nest distribution relation vector is input into a multi-agent depth deterministic strategy gradient algorithm, unmanned aerial vehicle path planning spatio-temporal data are discriminated and analyzed by means of a cooperative competition discrimination branch, complex relations such as nest layout and unmanned aerial vehicle paths can be fully considered, and resource allocation is optimized. According to the task indication information, a state vector is generated and a scheduling viewpoint is optimized, and then an autonomous cooperative inspection strategy is generated, so that the cooperative inspection efficiency of multiple unmanned aerial vehicles can be greatly improved, task conflicts are avoided, accurate matching of inspection task paths and task requirements is ensured, and a target area is effectively covered. The labor cost is reduced; and the accuracy and timeliness of routing inspection are improved.

Description

Technical Field

[0001] The embodiments of the present application belong to the field of drone technology, and specifically relate to an autonomous inspection method and system for drones with multi-nest take-off and landing in different locations. Background Art

[0002] Satellite remote sensing remote monitoring has the characteristics of wide coverage and strong macro information acquisition capabilities, and can conduct overall observations of large areas. "Space + low-altitude combination" means using the macro monitoring advantages of satellite remote sensing to provide global information for low-altitude drone inspections, and guide drones to plan inspection tasks more reasonably. However, in traditional inspection work, there are many technical problems. For example, in inspection tasks in large areas or complex environments, due to the lack of accurate analysis of the inspection space area, it is difficult to determine a reasonable machine nest layout, resulting in unreasonable setting of drone take-off points, affecting inspection efficiency. Moreover, for the collaborative inspection tasks of multiple drones, there is a lack of effective scheduling mechanism, and the drone path cannot be flexibly planned according to the actual machine nest distribution and task requirements. This makes it easy to have task conflicts, waste of resources, and incomplete inspection coverage during the inspection process. At the same time, there is a lot of manual intervention under the traditional method, which not only increases labor costs, but also affects the accuracy and timeliness of inspections due to human errors. Summary of the invention

[0003] The embodiments of the present application provide a method and system for autonomous inspection of drones with multi-nest take-off and landing at different locations, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0004] The embodiment of the present application provides an autonomous inspection method for unmanned aerial vehicles with multi-machine nests taking off and landing in different locations, which is applied to an autonomous inspection system for unmanned aerial vehicles. The method includes: identifying a target inspection space area, obtaining a multi-machine nest distribution relationship vector corresponding to the target inspection space area, and obtaining target inspection task indication information of the target inspection space area; the target inspection task indication information includes a collaborative inspection task requirement vector of multiple unmanned aerial vehicles; inputting the multi-machine nest distribution relationship vector into a multi-agent deep deterministic policy gradient algorithm, the multi-agent deep deterministic policy gradient algorithm includes a collaborative competition discrimination branch for discriminant analysis of unmanned aerial vehicle path planning spatiotemporal data based on the multi-machine nest distribution relationship vector; generating a first unmanned aerial vehicle cluster state vector corresponding to the collaborative competition discrimination branch according to the target inspection task indication information, optimizing a first collaborative inspection scheduling viewpoint of the collaborative competition discrimination branch according to the first unmanned aerial vehicle cluster state vector, obtaining a second collaborative inspection scheduling viewpoint, and generating an autonomous collaborative inspection strategy for the target inspection space area according to the second collaborative inspection scheduling viewpoint, so that the autonomous collaborative inspection strategy contains an inspection task path corresponding to the target inspection task indication information.

[0005] An embodiment of the present application provides an unmanned aerial vehicle autonomous inspection system, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.

[0006] An embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0007] The embodiments of the present application achieve efficient autonomous inspections by drones through a series of processes. First, the identification of the target inspection space area and the acquisition of relevant vectors and information enable the inspection task planning to accurately meet actual needs. The multi-machine nest distribution relationship vector is input into the multi-agent deep deterministic policy gradient algorithm, and the spatiotemporal data of drone path planning is discriminated and analyzed with the help of its collaborative competition discrimination branch, which can fully consider the complex relationships such as the machine nest layout and the drone path, and optimize resource allocation. According to the task instruction information, the state vector is generated and the scheduling viewpoint is optimized, and then an autonomous collaborative inspection strategy is generated. This can greatly improve the efficiency of collaborative inspections of multiple drones, avoid task conflicts, ensure that the inspection task path is accurately matched with the task requirements, effectively cover the target area, reduce labor costs and improve the accuracy and timeliness of inspections. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A flow chart of a method for autonomous inspection of unmanned aerial vehicles with multi-nest take-off and landing in different locations provided in an embodiment of the present application.

[0009] Figure 2 A schematic diagram of the structure of an autonomous drone inspection system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of them. Based on the embodiments in the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.

[0011] Figure 1 A method for autonomous inspection of unmanned aerial vehicles (UAVs) with multi-machine nests and remote take-off and landing is shown, which is applied to an autonomous inspection system of unmanned aerial vehicles. The method includes the following S102-S106.

[0012] S102. Identify the target inspection space area, obtain the multi-machine nest distribution relationship vector corresponding to the target inspection space area, and obtain the target inspection task indication information of the target inspection space area; the target inspection task indication information includes the collaborative inspection task requirement vector of multiple drones.

[0013] S104. Input the multi-machine-nest distribution relationship vector into a multi-agent deep deterministic policy gradient algorithm, wherein the multi-agent deep deterministic policy gradient algorithm includes a collaborative competitive discriminant branch for performing discriminant analysis on the UAV path planning spatiotemporal data based on the multi-machine-nest distribution relationship vector.

[0014] S106. Generate a first UAV cluster state vector corresponding to the collaborative competition judgment branch based on the target inspection task indication information, and optimize the first collaborative inspection scheduling viewpoint of the collaborative competition judgment branch based on the first UAV cluster state vector to obtain a second collaborative inspection scheduling viewpoint, and generate an autonomous collaborative inspection strategy for the target inspection space area based on the second collaborative inspection scheduling viewpoint, so that the autonomous collaborative inspection strategy contains an inspection task path corresponding to the target inspection task indication information.

[0015] In S102, the identification of the target inspection space area by the autonomous inspection system of the drone is the starting step of the entire inspection process. This identification process aims to accurately determine the spatial scope that needs to be inspected. For example, when inspecting a large industrial park, the area may contain many factories, warehouses, pipeline facilities, etc., and its spatial layout is complex and diverse. By using advanced geographic information system (GIS) technology combined with image recognition technology, the boundaries, topography, obstacle distribution and other information of the target inspection space area can be accurately outlined.

[0016] After identifying the target inspection space area, it is necessary to analyze the distribution relationship of multiple nests in the area. Nests are the take-off and landing points of drones, and their distribution directly affects the inspection efficiency and coverage. For example, in a mountainous power inspection area with an area of ​​100 square kilometers, 5 nests are arranged. The distribution of nests may be affected by factors such as terrain and distribution of power facilities. For example, some nests are located near substations in the valley, while others are located near communication base stations on the top of the mountain.

[0017] The coordinate position of each nest is obtained through an accurate positioning system (such as GPS or Beidou positioning system), and the terrain data of the target inspection space area and the location information of other related facilities are combined to construct a multi-nest distribution relationship vector. This vector can contain information such as the distance relationship between nests, relative height difference, and distance to key facilities (such as power towers, pumping stations, etc.) in the target inspection area. For example, the distance between nest A and nest B is 5 kilometers, the height difference between nest A and nest B is 200 meters, and the distance between nest A and the nearest power tower is 1 kilometer.

[0018] Among them, the target inspection task indication information is of guiding significance for the entire inspection task. The collaborative inspection task demand vector of multiple drones clarifies the various task requirements of drones during the inspection process. Taking the oil pipeline inspection as an example, if there are 100 kilometers of oil pipelines that need to be inspected in the target inspection space area, and there are 10 pumping stations that need to monitor the equipment status. Then the collaborative inspection task demand vector may include information such as the length of the pipeline that each drone needs to inspect, the type of pumping station equipment that needs to be monitored, and the time interval of the inspection. For example, drone 1 needs to inspect 30 kilometers of pipelines, focusing on monitoring the pressure equipment of 3 pumping stations, and conduct an inspection every 24 hours; drone 2 needs to inspect another 70 kilometers of pipelines, focusing on monitoring the flow equipment of 7 pumping stations, and conduct an inspection every 12 hours.

[0019] Next, S104 focuses on the application of the multi-agent deep deterministic policy gradient algorithm. The multi-agent deep deterministic policy gradient (MADDPG) algorithm is an algorithm suitable for decision optimization of multi-agent systems. In the drone autonomous inspection system, each drone can be regarded as an agent. The MADDPG algorithm continuously learns the interaction between the agent and the environment and optimizes the agent's strategy to achieve efficient collaborative inspection between multiple drones.

[0020] When the multi-dock distribution relationship vector is input into the cooperative competition discriminant branch in the MADDPG algorithm, the branch begins to discriminate the spatiotemporal data of the drone path planning. Taking an urban transportation facility inspection scene with 3 drone nests and 10 drones as an example, the spatiotemporal data includes the flight trajectory, speed, altitude and other information of the drone after taking off from the drone nest at different time points.

[0021] The collaborative competition judgment branch will analyze the collaborative and competitive relationship of drones in the path planning process based on the distribution relationship vectors of the nests. For example, if nests 1 and 2 are close to each other, and both have drones planning to take off to inspect traffic lights in the same area, then there is a competitive relationship, and it is necessary to determine how to allocate tasks to improve efficiency. If nest 3 is far away, but can cooperate with drones in nests 1 and 2 to cover traffic facilities in different directions, this reflects a collaborative relationship.

[0022] This branch will consider the impact of factors such as the distance and relative position between the nests on the drone path planning. For example, when the distance between the nests is within 5 kilometers, the flight path planning of the drone may need to consider more about avoiding collisions and reasonably allocating inspection areas; when the distance between the nests exceeds 10 kilometers, it may need to consider more about how to optimize energy consumption during long-distance flights.

[0023] In the final S106, the first drone cluster state vector corresponding to the collaborative competition discrimination branch is generated according to the target inspection task indication information. In a large wind farm inspection task, the target inspection task indication information includes inspection tasks for blades, towers, cabins and other equipment of 100 wind turbines. The first drone cluster state vector may include the number of drones (e.g., 8), the current task status of each drone (e.g., drone 1 is inspecting the blades of the 10th wind turbine, drone 2 is on the way to the 15th wind turbine, etc.), the remaining power of the drone (e.g., drone 1 has a remaining power of 80%, drone 2 has a remaining power of 70%, etc.), and other information related to the target inspection task (e.g., the number of wind turbines that have completed the inspection is 30, etc.).

[0024] Then, the first collaborative inspection scheduling viewpoint of the collaborative competition discrimination branch is optimized based on the first drone cluster state vector. In the above wind farm inspection scenario, if it is found that the inspection efficiency of some wind turbines is low, it may be because the drone's path planning is unreasonable or the distance between the nest and the area causes the drone to arrive too long. By analyzing the information in the first drone cluster state vector, such as the location and remaining power of the drone, the first collaborative inspection scheduling viewpoint can be optimized.

[0025] For example, the original first coordinated inspection scheduling concept may be to inspect the wind turbines in the order of their numbers, but the optimized concept may be adjusted to prioritize wind turbines that are close to the drone and have sufficient power based on the drone's location and power status, while avoiding the waste of resources caused by multiple drones inspecting the same area at the same time.

[0026] It can be understood that the second collaborative inspection scheduling viewpoint obtained after optimization is more in line with the current inspection task requirements and the actual status of the drone cluster. In a complex urban comprehensive facility inspection scenario, it includes inspections of facilities such as high-rise buildings, bridges, and underground pipelines. The second collaborative inspection scheduling viewpoint may comprehensively consider factors such as the priority of different facilities, the performance differences of drones (such as some drones are suitable for high-altitude inspections, while others are suitable for low-altitude close-up inspections), and the distance relationship between the machine nest and each facility. For example, for bridge inspection tasks in traffic congestion areas, the second collaborative inspection scheduling viewpoint will arrange drones with fast flight capabilities and sufficient power to execute first, while coordinating other drones to inspect the exterior walls of surrounding high-rise buildings to improve overall inspection efficiency.

[0027] When it comes to the generation and processing of the autonomous collaborative inspection strategy, the autonomous collaborative inspection strategy of the target inspection space area can be generated based on the second collaborative inspection scheduling viewpoint, so that the strategy contains the inspection task path corresponding to the target inspection task indication information. In a large chemical park inspection task, the target inspection task indication information requires regular safety inspections of various chemical equipment, storage tanks, pipelines, etc.

[0028] According to the second collaborative inspection scheduling viewpoint, the autonomous collaborative inspection strategy will plan a specific inspection task path for each drone. For example, the inspection task path of drone 1 is to take off from nest A, first inspect the pipeline in the north of the park, then check the level gauge and other equipment of storage tank No. 3, then fly to the reactor area in the middle of the park for inspection, and finally return to nest A. Drone 2 takes off from nest B and is responsible for inspecting equipment and pipelines in the south of the park. Through such an autonomous collaborative inspection strategy, multiple drones can efficiently and orderly complete the inspection tasks of the entire chemical park to ensure safe production in the chemical park.

[0029] Through the above detailed steps, the UAV autonomous inspection system can generate efficient autonomous collaborative inspection strategies based on the characteristics of the target inspection space area, the distribution relationship of the drone nests, and the target inspection task instruction information, using the multi-agent deep deterministic policy gradient algorithm, thereby achieving comprehensive, accurate, and efficient inspection tasks for the target inspection space area.

[0030] It is worth mentioning that the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm plays a vital role in the technical solution of the above-mentioned UAV autonomous inspection system. Its application will be elaborated in detail below.

[0031] In the UAV autonomous inspection system, there are multiple UAVs working together to complete the inspection task in the target inspection space area. Each UAV can be regarded as an intelligent agent, and these agents need to cooperate with each other and reasonably compete for resources in a complex environment to achieve efficient inspection goals. The MADDPG algorithm provides an effective decision optimization framework for this multi-agent system.

[0032] First, the multi-machine nest distribution relationship vector obtained from the target inspection spatial area identification (S102) stage provides basic information for the application of the MADDPG algorithm. The multi-machine nest distribution relationship vector contains various relationship information between the machine nests, such as the geographical location of the machine nests, the distance between the machine nests, and the location of the machine nests relative to the key facilities in the target inspection area. This information is of great significance in the MADDPG algorithm, because the machine nests are the take-off and landing points of the drones, and the distribution of the machine nests directly affects the path planning and task allocation of the drones.

[0033] Take an actual large-scale logistics park inspection scenario as an example. The logistics park is vast and contains numerous warehouses, transportation channels, sorting facilities, etc. For example, multiple machine nests are arranged in this logistics park, and the distribution of these machine nests is related to the layout of facilities in the park. Some machine nests are close to large warehouses to facilitate inspections of loading and unloading equipment and cargo stacking around the warehouses; some machine nests are close to transportation channels to facilitate monitoring of the operating conditions of logistics vehicles. When the MADDPG algorithm receives this multi-machine nest distribution relationship vector, it can convert this spatial relationship information into a data structure that can be processed within the algorithm, providing a basis for subsequent decision analysis.

[0034] At the same time, the collaborative inspection task demand vector of multiple drones in the target inspection task indication information is also closely related to the MADDPG algorithm. This demand vector clarifies the specific requirements of each drone in the inspection task. For example, in the inspection of the logistics park, different drones may be assigned different tasks. Some drones need to focus on inspecting the storage status of goods inside the warehouse, and need to check the quantity, type, and damage of the goods; some drones are responsible for inspecting the traffic conditions of the transportation channel, including vehicle flow, traffic jams, and whether the transportation facilities are operating normally. These different task requirements are regarded as different goals in the MADDPG algorithm, and the algorithm needs to optimize the behavior strategy of the drone based on these goals and information such as the distribution of the machine nest.

[0035] When the multi-drone distribution relationship vector is input into the MADDPG algorithm (S104), the collaborative competition discrimination branch in the MADDPG algorithm begins to play a role. This branch is specifically used to discriminate and analyze the spatiotemporal data of drone path planning based on the multi-drone distribution relationship vector. In the scenario of multi-drone collaborative inspection, drone path planning is a complex problem. It is necessary to consider not only how to efficiently cover the target inspection area, but also the collaborative and competitive relationship between drones.

[0036] For example, continuing with the logistics park inspection, multiple drones may take off from different nests for inspection at the same time. If the mission areas of two drones overlap, there will be competition, and they may compete for inspection resources in the same area, such as the best inspection angle, the shortest inspection path, etc. On the other hand, if the mission areas of the two drones are adjacent, they can improve the inspection efficiency through collaboration. For example, after inspecting its own area, a drone can pass some information related to the adjacent area to another drone to help it better plan the inspection path. The collaborative competition discrimination branch of the MADDPG algorithm can accurately identify these collaborative and competitive relationships by analyzing the nest distribution relationship vector.

[0037] In terms of spatiotemporal data for path planning, the MADDPG algorithm takes into account information such as the position, speed, and flight direction of the drone at different points in time. For example, during a logistics park inspection, a drone is located above the warehouse area at a certain point in time, with a speed of 10 meters per second and a flight direction toward the transportation channel. The algorithm will determine whether the current path of the drone is reasonable based on these spatiotemporal data as well as information such as the distribution of the drone nests and mission requirements. If it is unreasonable, for example, the drone is far away from the next target facility to be inspected, and there are other drones nearby that can reach the target facility more efficiently, then the algorithm will adjust the path planning of the drone through a collaborative competitive discrimination branch.

[0038] The MADDPG algorithm uses a deep neural network structure in this process. It contains a policy network and a value network. The policy network is used to generate the action strategy of the drone based on the current state (such as the distribution relationship of the drone nests, the current state of the drone, the target inspection task requirements, etc.). This action strategy includes the flight direction of the drone, speed adjustment, the selection of the next inspection target, etc. The value network is used to evaluate the quality of the action strategy generated by the policy network. For example, in the logistics park inspection scenario, the value network will evaluate the action strategy generated by the policy network based on factors such as whether the drone can efficiently complete the inspection task, whether it avoids collisions with other drones, and whether the drone nest resources are reasonably used.

[0039] In the process of generating the first drone cluster state vector (S106) corresponding to the collaborative competition discrimination branch based on the target inspection task indication information, the MADDPG algorithm plays a key role again. The first drone cluster state vector contains various information about the drone cluster, such as the number of drones, the current mission status of each drone, the remaining power, etc. The MADDPG algorithm accurately constructs this state vector by combining the previously received information (multi-machine nest distribution relationship vector, target inspection task indication information) and the experience accumulated in the previous decision-making process.

[0040] Take a more complex smart city inspection scenario as an example. There are various facilities in the city that need to be inspected, such as high-rise buildings, bridges, underground pipe networks, transportation hubs, etc. For example, if there are 10 drones participating in the inspection task, the MADDPG algorithm will construct the first drone cluster state vector based on the current task execution status of each drone (such as drone 1 is inspecting the facade of a high-rise building and has completed 50% of the inspection task; drone 2 is on the way to the bridge, etc.) and the remaining power of each drone (drone 1 has 60% remaining power, drone 2 has 70% remaining power, etc.). This vector can fully reflect the current state of the drone cluster and provide an accurate data basis for subsequent decision optimization.

[0041] In the process of optimizing the first collaborative inspection scheduling viewpoint of the cooperative competition discrimination branch according to the first drone cluster state vector and obtaining the second collaborative inspection scheduling viewpoint, the deep neural network structure of the MADDPG algorithm continuously learns and adjusts. In the smart city inspection scenario, if it is found that according to the first collaborative inspection scheduling viewpoint, the inspection efficiency of some areas is low, for example, the drone inspection density around some high-rise buildings is too high, and the inspection coverage of some underground pipe network areas is insufficient. The MADDPG algorithm will adjust the parameters of the policy network and the value network according to the information in the first drone cluster state vector.

[0042] For example, the algorithm may find that because some drone nests are far away from the underground pipe network area, the drone responsible for inspecting the area takes too long to fly, thereby reducing the inspection efficiency. Therefore, the algorithm adjusts the strategy network to change the drone's takeoff nest selection or adjust the flight path, so that the drone can reach the underground pipe network area for inspection more efficiently. At the same time, the value network will also re-evaluate the strategy based on the new strategy adjustment to ensure that the optimized scheduling view can truly improve the inspection efficiency.

[0043] Finally, when generating the autonomous collaborative inspection strategy for the target inspection space area based on the second collaborative inspection scheduling viewpoint, the MADDPG algorithm ensures that the strategy contains the inspection task path corresponding to the target inspection task indication information. In the smart city inspection scenario, for each facility that needs to be inspected, such as high-rise buildings, bridges, underground pipe networks, etc., the MADDPG algorithm will plan a specific inspection task path for each drone based on the second collaborative inspection scheduling viewpoint.

[0044] For example, UAV 1 may be planned to take off from a nest near the city center, first inspect the facades and rooftop facilities of several nearby high-rise buildings, then fly to the bridge area along a specific route to check the structural safety of the bridge, and finally go to a certain entrance of the underground pipe network to inspect a specific area of ​​the underground pipe network before returning to the nest. Such an autonomous collaborative inspection strategy can make full use of nest resources, consider the collaborative and competitive relationship between UAVs, and efficiently complete the inspection tasks in the complex environment of the smart city.

[0045] The MADDPG algorithm runs through the entire technical solution of the UAV autonomous inspection system. By processing and analyzing various data such as multi-machine nest distribution relationship vectors and target inspection task indication information, the behavior strategy of the UAV is continuously optimized to achieve efficient and accurate autonomous collaborative inspection tasks. This algorithm has unique advantages in dealing with complex collaboration and competition relationships in multi-agent systems and can adapt to different inspection scenarios, whether it is a large logistics park, smart city or other complex target inspection space area.

[0046] It can be seen that the embodiment of the present application realizes efficient autonomous inspection of drones through a series of processes. First, the identification of the target inspection space area and the acquisition of relevant vectors and information enable the inspection task planning to accurately meet actual needs. The multi-machine nest distribution relationship vector is input into the multi-agent deep deterministic policy gradient algorithm, and the spatiotemporal data discrimination analysis of drone path planning is carried out with the help of its collaborative competition discrimination branch, which can fully consider the complex relationships such as machine nest layout and drone path, and optimize resource allocation. According to the task indication information, the state vector is generated and the scheduling viewpoint is optimized, and then an autonomous collaborative inspection strategy is generated. This can greatly improve the efficiency of collaborative inspections of multiple drones, avoid task conflicts, ensure that the inspection task path is accurately matched with the task requirements, effectively cover the target area, reduce labor costs and improve the accuracy and timeliness of inspections.

[0047] The following is an example of an application scenario in forest resource inspection that combines satellite remote sensing remote monitoring capabilities to form a space + low-altitude technical solution.

[0048] In a vast forest area, it is set as the target inspection space area. This forest is vast and has complex terrain, including different types of trees, diverse landforms (such as mountains, rivers, valleys, etc.) and rich ecosystems.

[0049] First, the target inspection space area of ​​the forest should be identified. Through satellite remote sensing technology, macro image data of the forest is obtained from space. These satellite remote sensing data contain information such as the boundary information of the forest, the general terrain contour, and the distribution of different vegetation coverage areas. At the same time, combined with ground measurement data and geographic information system (GIS) data, the various sub-areas inside the forest are accurately determined, such as concentrated planting areas of different tree species, ecologically sensitive areas near rivers, and high-risk areas prone to fire. Based on this information, the multi-machine nest distribution relationship vector corresponding to the target inspection space area is obtained. For example, five machine nests are set up in the edge area of ​​this forest. The distribution of the machine nests should take into account the coverage effect of the entire forest area and the relative position relationship with the special areas inside the forest. For example, some machine nests are close to the foot of the mountain, which is convenient for inspecting the forest area on the mountain; some machine nests are located near the estuary of the river, which can better monitor the situation of the forest on both sides of the river. The distance and relative orientation between the machine nests are accurately quantified and integrated into the multi-machine nest distribution relationship vector.

[0050] At the same time, the target inspection task instruction information of the target inspection space area is obtained. Forest inspection tasks are complex and diverse, and the collaborative inspection task requirement vectors of multiple drones in the target inspection task instruction information cover a lot of content. For example, it is necessary to locate and monitor the health of rare trees in the forest, which requires drones to have high-resolution image acquisition capabilities and be able to accurately identify tree species; it is also necessary to investigate forest fire hazards, which requires drones to be equipped with thermal imaging equipment and conduct inspections according to specific inspection routes and time intervals. In addition, ecological environment monitoring is required for wildlife habitats in forests, and drones must be able to collect environmental data (such as temperature, humidity, air quality, etc.).

[0051] Then, the multi-machine nest distribution relationship vector is input into the multi-agent deep deterministic policy gradient algorithm. The cooperative competition discrimination branch in the algorithm begins to discriminate and analyze the spatiotemporal data of the UAV path planning based on the multi-machine nest distribution relationship vector. In the forest inspection scenario, the spatiotemporal data of the UAV path planning is very complex. For example, considering the inspection priority of different areas in the forest, the flight path of the UAV after taking off from the machine nest should avoid obstacles such as mountains, and arrive at specific inspection areas at different time points. The cooperative competition discrimination branch analyzes the cooperative and competitive relationship of the UAV in the path planning process according to the distribution relationship vector of the machine nest. For example, if two machine nests are close to each other and both have UAVs responsible for inspecting the same high-risk fire area in the forest, there is a competitive relationship, and it is necessary to arrange the inspection time and tasks reasonably to avoid duplication of work; if some machine nests are responsible for a large inspection area, it is necessary to cooperate with the UAVs of other machine nests to share some key inspection data (such as real-time images of fire hazard points) to improve the overall inspection efficiency.

[0052] The first drone cluster state vector corresponding to the collaborative competition discrimination branch is generated according to the target inspection task indication information. This vector reflects the state of the drone cluster at the beginning of the inspection task. For example, there are 10 drones participating in the forest inspection task. The first drone cluster state vector contains the initial position of each drone (i.e., which nest it takes off from), the initial power, the status of the equipment carried (such as thermal imaging equipment, image acquisition equipment, etc.), and the initial task allocation (such as drone 1 is responsible for inspecting a rare tree planting area, etc.).

[0053] Then, the first collaborative inspection scheduling viewpoint of the collaborative competition discrimination branch is optimized according to the state vector of the first UAV cluster to obtain the second collaborative inspection scheduling viewpoint. During the forest inspection process, if a certain UAV is found to have a low initial battery or a temporary failure of the equipment it carries, the first collaborative inspection scheduling viewpoint may cause the UAV to fail to complete the task well. Through optimization, the second collaborative inspection scheduling viewpoint will readjust the task allocation. For example, the high-difficulty inspection tasks that the UAV was originally responsible for (such as the inspection of fire hazards in deep mountain areas) are adjusted to relatively simple tasks (such as monitoring wildlife activities on the edge of the forest), while allowing UAVs with sufficient power and good equipment status to undertake more important tasks.

[0054] Finally, the autonomous collaborative inspection strategy for the target inspection space area is generated based on the second collaborative inspection scheduling viewpoint, so that the autonomous collaborative inspection strategy contains the inspection task path corresponding to the target inspection task indication information. In the forest inspection, each drone performs the inspection according to this autonomous collaborative inspection strategy. For example, drone 1 takes off from the nest near the mountain, first flies along one side of the mountain, and uses thermal imaging equipment to check for forest fire hazards. During the inspection, if a suspected fire hazard point is found, the image data and location information will be sent to other drones and the ground control center in a timely manner; then it flies to the rare tree planting area according to the predetermined path, and switches to high-resolution image acquisition equipment to monitor the health of rare trees; drone 2 takes off from the nest near the river estuary, flies along both sides of the river, collects environmental data and monitors the ecological environment of wildlife habitats, and cooperates with drone 1 to complete the inspection task of the entire forest area.

[0055] Satellite remote sensing remote monitoring always plays an important role in the entire inspection process. Satellites regularly obtain macroscopic image data of forests, which can be used to update the information of the target inspection space area. For example, if the vegetation coverage in a certain area of ​​the forest is found to have changed abnormally (possibly due to pests or illegal logging, etc.), satellite remote sensing data can promptly transmit this information to the ground control center. The ground control center adjusts the inspection task instruction information and inspection strategy of the drone based on the macro information provided by satellite remote sensing. For example, when the satellite finds that the vegetation color in a certain forest area is abnormal and suspected of an outbreak of pests and diseases, the ground control center can adjust the inspection path of the drone, allowing the drone to give priority to detailed inspections of the area and increase the frequency of inspections. This combination of satellite remote sensing remote monitoring and drone low-altitude inspections achieves a perfect combination of space + low altitude, and conducts comprehensive, efficient and accurate inspections of forest resources.

[0056] In some examples, the obtaining of target inspection task indication information of the target inspection space area in S102 includes: locating the inspection task label from the target inspection space area, and obtaining the target inspection space sub-area containing the inspection task label; inputting the target inspection space sub-area and the associated space sub-area of ​​the target inspection space sub-area in the target inspection space area into the task priority parsing network to obtain the task priority of the parsed inspection task label; obtaining the machine nest take-off and landing time efficiency weight corresponding to the inspection task label; determining the target inspection task label in the located inspection task label based on the task priority and the machine nest take-off and landing time efficiency weight; and obtaining the target inspection task indication information of the target inspection task label.

[0057] Among them, obtaining the target inspection task indication information of the target inspection space area in S102 first locates the inspection task label from the target inspection space area and obtains the target inspection space sub-area containing the inspection task label. This process involves the detailed division and mark identification of the target inspection space area. For example, in a large industrial park inspection scenario, the target inspection space area is the entire industrial park, which contains many different types of facilities and areas, such as production workshops, storage areas, power facility areas, etc. Each area may have its own specific inspection task label. For example, the inspection task label of the production workshop may be equipment operation status monitoring, and the storage area may be cargo storage safety monitoring, etc.

[0058] After locating these inspection task labels, the target inspection space sub-area and the associated space sub-area of ​​the target inspection space sub-area in the target inspection space area are input into the task priority parsing network. The task priority parsing network is a network model used to analyze the importance of different tasks. For example, in the above industrial park, the equipment operation status monitoring in the production workshop may be parsed as a higher task priority in the task priority parsing network due to the high value, high risk and key impact of the equipment on the production process; while the priority of cargo storage safety monitoring in the storage area may be relatively lower, but if the storage area stores hazardous chemicals, its priority will be increased accordingly. This task priority parsing is based on a comprehensive consideration of many factors, including the value of the facilities in the sub-area, the probability of possible failure or danger, and the impact on the overall system. For example, the task priority of equipment operation status monitoring in the production workshop is parsed as 0.8 (the value range can be set according to the specific network, for example, 0-1 here means the priority is from low to high), and the task priority of cargo storage safety monitoring in the storage area is 0.6.

[0059] Next, obtain the machine nest take-off and landing time efficiency weight corresponding to the inspection task label. The machine nest take-off and landing time efficiency weight reflects the time efficiency impact of taking off from different machine nests to execute specific inspection task labels. For example, in an industrial park, if there is a machine nest close to the production workshop, then for the inspection task label of production workshop equipment operation status monitoring, the machine nest take-off and landing time efficiency weight of taking off from this machine nest to execute the task may be higher, such as 0.9 (also such as 0-1 The numerical range), because taking off from this machine nest can reach the target inspection area faster, reduce flight time, and improve inspection efficiency; and if there is a machine nest close to the storage area, for the task of cargo storage safety monitoring in the storage area, its machine nest take-off and landing time efficiency weight may be 0.8.

[0060] Then, according to the task priority and the weight of the time efficiency of the take-off and landing of the machine nest, the target inspection task label is determined in the located inspection task labels. Continuing with the industrial park as an example, each inspection task label is comprehensively evaluated through a specific calculation method (for example, it can be the product of the task priority and the weight of the time efficiency of the take-off and landing of the machine nest). For example, the comprehensive evaluation value of the equipment operation status monitoring in the production workshop is 0.8*0.9=0.72, and the comprehensive evaluation value of the cargo storage safety monitoring in the storage area is 0.6*0.8=0.48. In this way, the inspection task label with the highest comprehensive evaluation value is determined as the target inspection task label. In this example, the equipment operation status monitoring in the production workshop becomes the target inspection task label.

[0061] Finally, the target inspection task indication information of the target inspection task label is obtained. The target inspection task indication information includes the collaborative inspection task demand vector of multiple drones and other contents. For the target inspection task label of production workshop equipment operation status monitoring, the target inspection task indication information may include how many drones are required to participate in the inspection (for example, 3 drones), the specific inspection path planning of each drone (for example, drone 1 is responsible for the equipment in workshop area A, drone 2 is responsible for the equipment in workshop area B, and drone 3 is responsible for the equipment in workshop area C), the inspection time interval (for example, once every 2 hours), the type of data that needs to be collected (such as equipment temperature, vibration frequency, etc.), etc.

[0062] In this way, by accurately locating the inspection task label and obtaining the target inspection task indication information, the drone inspection task can be planned more scientifically and reasonably. Taking the industrial park as an example, by determining the task priority and the weight of the take-off and landing time of the machine nest, drone resources can be allocated to the most critical, urgent and most timely inspection tasks from the machine nest. For example, when the equipment operation status monitoring task has a high priority and the take-off and landing time weight of the machine nest is high, it can ensure that the equipment in the production workshop is inspected in a timely and effective manner to avoid production losses caused by equipment failure. At the same time, according to the specific target inspection task indication information, multiple drones can perform inspections according to the planned collaborative inspection task demand vector, which improves the accuracy and efficiency of the inspection, reduces unnecessary flights and resource waste, and can ensure that key areas and tasks are focused on, ensuring the safety and normal operation of the entire inspection area.

[0063] In other examples, the obtaining of target inspection task indication information of the target inspection space area in S102 includes: dividing the target inspection space area to obtain inspection space sub-areas; inputting the divided inspection space sub-areas and set reference features into a task feature mining network, so that the task feature mining network generates an inspection task unique hot vector based on the inspection space sub-areas, wherein the set reference features are used to characterize that the task feature mining network generates the inspection task unique hot vector based on the inspection space sub-areas; and performing residual connection on the inspection task unique hot vector through a deep residual connection network to obtain the target inspection task indication information.

[0064] It can be understood that the deep residual connection network is debugged through the following steps: obtaining a first a priori debugging example set, wherein the first a priori debugging example set includes a plurality of past inspection area task tuples, each of the past inspection area task tuples includes mutually corresponding past inspection space areas and past inspection task unique hot vectors; performing residual connection on each of the past inspection space areas through an initial convolutional neural network, and performing residual connection on each of the past inspection task unique hot vectors through the deep residual connection network to obtain a plurality of past inspection space area vectors and past inspection task unique hot vectors; identifying each of the past inspection space areas; performing residual connection on each of the past inspection space areas; and performing residual connection on each of the past inspection task unique hot vectors through the deep residual connection network to obtain a plurality of past inspection space area vectors and past inspection task unique hot vectors; and identifying each of the past inspection space areas; The first correlation coefficient between the past inspection space area vector and each of the past inspection task unique hot vectors is determined, and the second correlation coefficient between each of the past inspection space area vectors and each of the past inspection task unique hot vectors is determined according to the adaptation characteristics between the past inspection space area and the past inspection task unique hot vector; a first network debugging error is determined according to the first correlation coefficient and the second correlation coefficient, and the deep residual connection network and the initial convolutional neural network are optimized according to the first network debugging error until the first network debugging error is less than the first debugging error threshold, and the debugging is terminated.

[0065] In another preferred embodiment, after performing residual connection on the inspection task unique hot vector through the deep residual connection network to obtain the target inspection task indication information, the method also includes: acquiring a first policy construction reference feature and a second policy construction reference feature, and mapping the first policy construction reference feature and the second policy construction reference feature into a corresponding policy task matching vector; wherein the first policy construction reference feature is used to characterize the first template inspection task label that the multi-agent deep deterministic policy gradient algorithm needs to correspond to when generating the autonomous collaborative inspection strategy, and the second policy construction reference feature is used to characterize the second template inspection task label that needs to be excluded when generating the autonomous collaborative inspection strategy; the policy task matching vector is integrated into the target inspection task indication information to adjust the target inspection task indication information.

[0066] In this technical solution, obtaining the target inspection task instruction information of the target inspection space area in S102 includes a series of rigorous operation procedures.

[0067] First, the target inspection space area is divided into inspection space sub-areas. This division process needs to be based on specific rules or standards, such as division by geographical area, functional area, etc. Taking the inspection of a large power facility as an example, the target inspection space area is the entire power facility coverage area, which can be divided into different inspection space sub-areas according to different functions such as substations, transmission lines, and distribution equipment.

[0068] Next, the divided inspection space sub-areas and set reference features are input into the task feature mining network. The set reference features here are used to characterize the basis for the task feature mining network to generate the inspection task unique hot vector according to the inspection space sub-areas. The task feature mining network is a network model specifically used to mine task features. It can generate the inspection task unique hot vector according to the input inspection space sub-areas and set reference features. For example, in the power facility inspection scenario, for example, for a certain transmission line sub-area, the task feature mining network generates an inspection task unique hot vector [0, 1, 0, 0, ...] according to the line length, voltage level, surrounding environment and other factors of the sub-area (these factors may be included in the set reference features), where each element represents a possible inspection task feature, such as 1 may represent the unique hot encoding of the insulator inspection task in the sub-area.

[0069] Then, the inspection task one-hot vector is residually connected through the deep residual connection network to obtain the target inspection task indication information. The deep residual connection network is a special neural network structure that can effectively avoid the gradient vanishing or gradient exploding problems when processing information, thereby better retaining information features. The principle is to enable the network to learn the residual relationship between input and output by constructing residual blocks. For example, there is a residual block in the deep residual connection network, whose input is x, and the output after two convolutional layers (weights W1 and W2 respectively) is F(x), then the output of the residual block is H(x)=F(x)+x. Here, F(x) can be expressed as F(x)=W2*σ(W1*x), where σ is the activation function (such as ReLU function). This residual connection method helps the network better learn the features in the one-hot vector of the inspection task, thereby generating the target inspection task indication information.

[0070] Regarding the debugging of the deep residual connection network, a first a priori debugging example set is obtained, which contains multiple past inspection area task tuples, each of which includes mutually corresponding past inspection space areas and past inspection task unique hot vectors. For example, the past inspection area may be the inspection area of ​​different time periods or different power facilities in the past, and each area has a corresponding inspection task unique hot vector. Residual connections are made to each past inspection space area through the initial convolutional neural network, and residual connections are made to each past inspection task unique hot vector through the deep residual connection network, so as to obtain multiple past inspection space area vectors and past inspection task unique hot vectors.

[0071] Next, identify the first correlation coefficient between each past inspection space area vector and each past inspection task unique hot vector. Here, the Pearson correlation coefficient can be used to calculate the correlation. For two variables X and Y, the Pearson correlation coefficient formula is: ρ(X, Y) = Cov(X, Y) / (σ(X)*σ(Y)), where Cov(X, Y) is the covariance of X and Y, and σ(X) and σ(Y) are the standard deviations of X and Y, respectively. This coefficient is used to measure the degree of linear correlation between two vectors. At the same time, the second correlation coefficient between each past inspection space area vector and each past inspection task unique hot vector is determined based on the adaptation characteristics between the past inspection space area and the past inspection task unique hot vector. This adaptation feature may involve factors such as the matching relationship between the special attributes of the inspection area and the task requirements.

[0072] The first network debugging error is determined based on the first correlation coefficient and the second correlation coefficient. For example, the first network debugging error E can be calculated by a custom formula, such as E=α*(1-ρ1)+β*(1-ρ2), where ρ1 is the first correlation coefficient, ρ2 is the second correlation coefficient, and α and β are weight coefficients used to adjust the importance of the two correlation coefficients in the error calculation. Based on this first network debugging error, the deep residual connection network and the initial convolutional neural network are optimized until the first network debugging error is less than the first debugging error threshold, and the debugging is terminated. This process continuously adjusts the parameters of the network to improve the accuracy of the network's generation of inspection task indication information.

[0073] In another preferred embodiment, after residual connection is performed on the inspection task unique hot vector through the deep residual connection network to obtain the target inspection task indication information, it also includes obtaining the first strategy construction reference feature and the second strategy construction reference feature, and mapping them into corresponding strategy task matching vectors. For example, in the inspection of power facilities, the first strategy construction reference feature characterizes the first template inspection task label that the multi-agent deep deterministic policy gradient algorithm needs to correspond to in the generated autonomous collaborative inspection strategy, such as the inspection task label for a specific voltage level transmission line; the second strategy construction reference feature characterizes the second template inspection task label that the multi-agent deep deterministic policy gradient algorithm needs to exclude in the generated autonomous collaborative inspection strategy, such as the task label corresponding to the equipment that has been overhauled and does not need to be inspected again in the short term. The strategy task matching vector is integrated into the target inspection task indication information to adjust the target inspection task indication information. This process can optimize the target inspection task indication information according to specific strategy requirements to make it more in line with actual inspection strategy requirements.

[0074] In this way, by dividing the target inspection space area in detail and using the task feature mining network and the deep residual connection network to generate the target inspection task indication information, the task characteristics of different inspection space sub-areas can be accurately mined. Taking the inspection of power facilities as an example, the inspection task requirements of different functional areas (such as substations, transmission lines, etc.) can be accurately reflected in the target inspection task indication information. The debugging process of the deep residual connection network calculates the network debugging error by using past inspection data and correlation coefficients, continuously optimizes the network, and improves the accuracy of the generation of target inspection task indication information. In addition, by integrating the strategy task matching vector into the target inspection task indication information, it can be adjusted according to the strategy requirements of the multi-agent deep deterministic policy gradient algorithm, further optimize the inspection task planning, improve the inspection efficiency, avoid invalid operations on areas that do not need to be inspected, and ensure the focus on key inspection tasks, ensuring the effectiveness and accuracy of the entire inspection task.

[0075] Under a preferred design idea, the discriminant analysis of the UAV path planning spatiotemporal data based on the multi-machine-nest distribution relationship vector in S104 includes: in the collaborative competition discrimination branch, determining the path planning transient spatiotemporal data according to the UAV path planning spatiotemporal data; generating a spatiotemporal feature relationship network according to the path planning transient spatiotemporal data, and generating a task identification feature relationship network and an inspection point feature relationship network according to the multi-machine-nest distribution relationship vector; performing a feature collision operation based on the spatiotemporal feature relationship network, the task identification feature relationship network and the inspection point feature relationship network, and determining the first collaborative inspection scheduling viewpoint corresponding to the collaborative competition discrimination branch according to the result of the feature collision operation.

[0076] Based on this design idea, first, in the collaborative competition judgment branch, the transient spatiotemporal data of path planning is determined based on the spatiotemporal data of the UAV path planning. The spatiotemporal data of the UAV path planning contains a lot of information, such as the position coordinates of the UAV at different times (which can be expressed by three-dimensional coordinates (x, y, z), x and y represent the plane position, and z represents the height), flight speed (in meters per second, such as 10 meters per second), flight direction (expressed in angles, such as 0°-360°), etc. The transient spatiotemporal data of path planning selects data in a specific state or a specific time period from these complete spatiotemporal data. For example, for a group of UAVs that are performing inspection tasks, the UAV path planning data within a specific time period (such as 10:00-10:10) is selected as the transient spatiotemporal data of path planning. During this time period, the UAV may be on the way from the machine nest to the target inspection area, or is inspecting a key inspection point.

[0077] Next, a spatiotemporal feature relationship network is generated based on the transient spatiotemporal data of path planning. This spatiotemporal feature relationship network is a data structure used to describe the relationship between various features in the transient spatiotemporal data of UAV path planning. It can be represented by a graph structure in graph theory, where nodes can represent the state of the UAV at different times (position, speed, direction, etc.), and edges represent the relationship between different states, such as the transition relationship between states at adjacent times. When constructing this spatiotemporal feature relationship network, an adjacency matrix can be used to represent the connection relationship between nodes. For example, if a UAV has n different states in the transient spatiotemporal data of path planning, then the adjacency matrix A is an n*n matrix, where the element Aij indicates whether there is a connection relationship from state i to state j. If so, Aij=1, otherwise Aij=0.

[0078] Then, the task identification feature relationship network and the inspection point feature relationship network are generated based on the multi-machine nest distribution relationship vector. The multi-machine nest distribution relationship vector contains information such as the distance between the machine nests (in meters, such as the distance between machine nest A and machine nest B is 500 meters), the orientation of the machine nest relative to the inspection area (expressed in angles, such as machine nest C is located in the northeast of the inspection area, with an angle of 45°), etc. The task identification feature relationship network is mainly used to describe the relationship between different tasks and machine nests, such as the execution priority or feasibility of different types of inspection tasks (such as equipment appearance inspection, equipment performance testing, etc.) in different machine nests. The inspection point feature relationship network focuses on describing the relationship between inspection points (such as specific power towers, specific nodes of oil pipelines, etc.) and machine nests, including the distance from the machine nest to the inspection point, the best path to the inspection point, etc. When constructing the task identification feature relationship network and the inspection point feature relationship network, the relationship between features can also be expressed in the form of a graph structure or a matrix.

[0079] Next, feature collision operations are performed based on the spatiotemporal feature network, the task identification feature network, and the inspection point feature network. The feature collision operation here is an operation that matches, compares, and analyzes features in different networks. For example, feature collision can be achieved by calculating the distance measurement between the same or similar features in different networks. For example, in the spatiotemporal feature network, there is a feature v1 that represents the speed of the drone at a certain moment, and in the task identification feature network, there is a feature v2 that represents the speed requirement of the drone for a certain task. The degree of collision between the two features is measured by calculating the value of |v1-v2|. If this value is small, it means that the current speed of the drone is more consistent with the speed requirement of the task; if this value is large, it means that there is a large mismatch.

[0080] For the comprehensive feature collision operation between multiple features, a weighted summation method can be used. For example, the feature collision metric C for n features can be expressed as C = Σ(wi*|fi1-fi2|), where wi is the weight of the i-th feature, reflecting the importance of the feature in the overall feature collision metric, fi1 is the i-th feature in the spatiotemporal feature relationship network, and fi2 is the corresponding feature in the task identification feature relationship network or the inspection point feature relationship network.

[0081] Finally, the first collaborative inspection scheduling viewpoint corresponding to the collaborative competition discrimination branch is determined based on the result of the feature collision operation. The first collaborative inspection scheduling viewpoint is a strategic viewpoint on how to arrange multiple drones to collaboratively perform inspection tasks. For example, if the result of the feature collision operation shows that the current flight path and speed of a drone are greatly mismatched with the mission requirements and the needs of the inspection points, then the flight mission of the drone may be adjusted in the first collaborative inspection scheduling viewpoint, such as changing its flight path to make it closer to the optimal path, or adjusting its flight speed to meet the time requirements of the mission. At the same time, if it is found that multiple drones have task competition at certain inspection points (such as multiple drones approaching a high-priority inspection point at the same time), the first collaborative inspection scheduling viewpoint will also coordinate, for example, according to factors such as the distance between the machine nest and the inspection point, the current state of the drone, etc., arrange one of the drones to perform the task of the inspection point first, and other drones adjust the task order or perform other related tasks.

[0082] It can be seen that by accurately determining the transient spatiotemporal data of path planning and constructing different feature relationship networks for feature collision operations, various factors in drone path planning can be comprehensively and deeply analyzed. Taking numerical examples as an example, by calculating the specific measurement values ​​between different features, such as the difference in speed features, the degree of match between drone path planning and task requirements can be accurately found. The first collaborative inspection scheduling viewpoint determined based on these analysis results can effectively coordinate the cooperation and competition relationship between multiple drones. Avoid task conflicts between drones and improve task execution efficiency. At the same time, ensure that the path planning and task arrangement of each drone match the actual situation such as the distribution of machine nests and inspection points, so as to ensure the smooth progress of the entire inspection task and improve the accuracy and effectiveness of the inspection.

[0083] Under another preferred design idea, the discriminant analysis of the UAV path planning spatiotemporal data based on the multi-machine-nest distribution relationship vector in S104 also includes: obtaining a first statistical value, and generating a number of initial inspection path spatiotemporal features corresponding to the first statistical value, and generating a number of UAV path planning spatiotemporal data corresponding to the first statistical value based on the initial inspection path spatiotemporal features; inputting the UAV path planning spatiotemporal data corresponding to the first statistical value into the multi-agent deep deterministic policy gradient algorithm respectively, and performing discriminant analysis on the UAV path planning spatiotemporal data corresponding to the first statistical value based on the multi-machine-nest distribution relationship vector to obtain a plurality of pending autonomous collaborative inspection strategies, the number of which is the same as the first statistical value; and determining an autonomous collaborative inspection strategy from the plurality of pending autonomous collaborative inspection strategies.

[0084] In this design concept, the discriminant analysis of the spatiotemporal data of the UAV path planning based on the multi-machine nest distribution relationship vector in S104 includes a further operation process.

[0085] First, obtain the first statistical value. This first statistical value is a certain value, which determines the amount of data in subsequent operations. For example, the first statistical value is 5, and the determination of this value may be based on various factors, such as computing resources, expected policy diversity, etc.

[0086] Next, generate the initial inspection path spatiotemporal features corresponding to the number of the first statistical value. The initial inspection path spatiotemporal features are the basis for constructing the spatiotemporal data of the UAV path planning. For example, for each initial inspection path spatiotemporal feature, it includes the starting position coordinates of the UAV (for example, (x0, y0, z0), where x0, y0 represent plane coordinates, z0 represents the height, in meters), the starting time t0 (in seconds, for example, t0 = 0 seconds), the initial speed v0 (such as 10 meters / second), the initial flight direction θ0 (in degrees, for example, θ0 = 0°), etc. These initial inspection path spatiotemporal features are generated by specific rules or algorithms to cover different possible situations.

[0087] Generate the UAV path planning spatiotemporal data corresponding to the first statistical value based on the initial inspection path spatiotemporal characteristics. Take one of the initial inspection path spatiotemporal characteristics as an example, and generate the UAV path planning spatiotemporal data according to a certain dynamic model. For example, using the uniform linear motion model (the actual situation will be more complicated), in the plane rectangular coordinate system, the position coordinates (x, y) of the UAV at time t can be calculated according to the formula x = x0 + v0 * cos (θ0) * t, y = y0 + v0 * sin (θ0) * t. At the same time, considering the change in the height direction (for example, rising or falling at a constant vertical speed vz), z = z0 + vz * t. Combine the information such as (x, y, z) at different times t, speed v (including the composite speed in the horizontal and vertical directions), and flight direction θ to obtain a complete UAV path planning spatiotemporal data. Since there are initial inspection path spatiotemporal characteristics corresponding to the first statistical value (here 5), 5 sets of such UAV path planning spatiotemporal data will be generated.

[0088] Then, the spatiotemporal data of the UAV path planning corresponding to the first statistical value are respectively input into the multi-agent deep deterministic policy gradient algorithm. The multi-agent deep deterministic policy gradient algorithm (MADDPG) is an algorithm for decision optimization of multi-agent systems. In this process, the algorithm takes the spatiotemporal data of each UAV path planning as the input state, considers the influence of the multi-machine nest distribution relationship vector, and performs decision optimization. The multi-machine nest distribution relationship vector contains the coordinates of the machine nest (for example, the coordinates of machine nest 1 are (x1, y1, z1), the coordinates of machine nest 2 are (x2, y2, z2), etc., in meters), the distance between the machine nests (such as the distance between machine nest 1 and machine nest 2 d12 = √((x2-x1)2+(y2-y1)2+(z2-z1)2)), the relative position relationship between the machine nest and the inspection area, and other information.

[0089] According to the multi-machine nest distribution relationship vector, discriminant analysis is performed on the UAV path planning spatiotemporal data corresponding to the first statistical value to obtain multiple pending autonomous collaborative inspection strategies, and the number of pending autonomous collaborative inspection strategies is the same as the first statistical value. In the discriminant analysis process, the MADDPG algorithm considers the relationship between various factors in the UAV path planning spatiotemporal data (such as the flight trajectory, speed, direction, etc. of the UAV) and the nest-related information in the multi-machine nest distribution relationship vector (such as the position of the nest, the distance between the nests, etc.). For example, according to the distance between the nest and the current position of the UAV, it is judged whether the UAV needs to adjust the flight direction to go to a closer nest for replenishment or adjust the task; according to the relative position of the nest and the inspection target, it is evaluated whether the current UAV path is the optimal path, etc. Through such discriminant analysis, for each set of input UAV path planning spatiotemporal data, a pending autonomous collaborative inspection strategy will be obtained. Since there are 5 sets of input data, 5 pending autonomous collaborative inspection strategies will be obtained.

[0090] Finally, an autonomous collaborative inspection strategy is determined from multiple pending autonomous collaborative inspection strategies. This determination process may be based on multiple evaluation criteria. For example, the task completion efficiency index of each pending autonomous collaborative inspection strategy can be calculated. For example, the task completion efficiency index E can be calculated by the formula E=Σ(ti*wi) / T, where ti is the time spent to complete each subtask (such as inspecting a specific area or equipment), wi is the weight of the subtask (reflecting the importance of the task), and T is the total time budget. The pending autonomous collaborative inspection strategy with the highest task completion efficiency index is selected as the final autonomous collaborative inspection strategy.

[0091] Therefore, by obtaining the first statistical value and generating a corresponding number of initial inspection path spatiotemporal features, a variety of different initial situations can be considered to increase the diversity of strategies. The spatiotemporal data of drone path planning generated based on these features provides rich input for the multi-agent deep deterministic policy gradient algorithm, so that different inspection strategies can be explored more comprehensively. By performing discriminant analysis on these data separately to obtain multiple pending autonomous collaborative inspection strategies, and then determining the final strategy from them, it is possible to find the optimal autonomous collaborative inspection strategy based on the distribution relationship vectors of multiple machine nests. This helps to improve the efficiency and accuracy of drone inspections, optimize drone path planning, fully consider the impact of the layout of machine nests on inspection tasks, avoid unnecessary flights and waste of resources, and ensure the efficient completion of inspection tasks.

[0092] In an alternative technical solution, the method of generating a first drone cluster state vector corresponding to the collaborative competition discrimination branch based on the target inspection task indication information in S106 includes: obtaining the original processing object of the first collaborative competition discrimination branch of the multi-agent deep deterministic policy gradient algorithm; performing feature embedding processing on the target inspection task indication information, and fusing the target inspection task indication information after feature embedding processing with the original processing object to obtain target inspection task fusion information; inputting the target inspection task fusion information into the corresponding drone tracking and monitoring algorithm, and performing feature embedding processing on the monitoring results of the drone tracking and monitoring algorithm to obtain the first drone cluster state vector.

[0093] In the above technical scheme, the debugging process of the drone tracking and monitoring algorithm includes: iterating the cooperative competition discrimination branch in the multi-agent deep deterministic policy gradient algorithm to obtain the iterated drone tracking and monitoring algorithm, and locking the algorithm variables of the multi-agent deep deterministic policy gradient algorithm, the drone tracking and monitoring algorithm includes long short-term memory modules corresponding to multiple cooperative competition discrimination branches of the multi-agent deep deterministic policy gradient algorithm; obtaining a second a priori debugging example set, the second a priori debugging example set includes at least one historical inspection record; extracting historical multi-agent scheduling strategies, historical inspection task indication information and historical inspection quantization vectors from the historical inspection records; implementing several rounds of perturbation configuration processing on the historical multi-agent scheduling strategy through the multi-agent deep deterministic policy gradient algorithm to obtain a multi-agent perturbation scheduling strategy, and caching each of the perturbation configuration processing configured in the first perturbation feature of the historical multi-agent scheduling strategy. Characteristic; input the historical inspection quantization vector into the multi-agent deep deterministic policy gradient algorithm, so that the multi-agent deep deterministic policy gradient algorithm implements several rounds of disturbance detection processing on the multi-agent disturbance scheduling strategy according to the historical inspection quantization vector; in each round of disturbance detection processing, input the historical inspection task indication information into the UAV tracking and monitoring algorithm, obtain the UAV cluster state vector sample corresponding to each collaborative competition discrimination branch, optimize the output of the corresponding collaborative competition discrimination branch according to the UAV cluster state vector sample, so as to obtain the second disturbance feature corresponding to the single round of disturbance detection processing, and several rounds of disturbance detection processing correspond to several rounds of disturbance configuration processing; determine the second network debugging error according to the second disturbance feature and the first disturbance feature of the corresponding disturbance configuration processing, and optimize the algorithm variables of each long short-term memory module according to the second network debugging error, until the UAV tracking and monitoring algorithm meets the first monitoring quality requirement, and obtain the UAV tracking and monitoring algorithm.

[0094] In detail, generating the first drone cluster state vector corresponding to the collaborative competition discrimination branch according to the target inspection task indication information in S106 involves a series of complex and interrelated operations.

[0095] First, the original processing object of the first cooperative competition discriminant branch of the multi-agent deep deterministic policy gradient algorithm is obtained. This original processing object is the initial basis for the discriminant branch to process, and it contains basic information related to the multi-agent system, such as the distribution relationship of multiple drone nests, the initial drone task allocation strategy and other basic data structures.

[0096] Next, feature embedding is performed on the target inspection task indication information. Feature embedding is a technical means of mapping various features in the target inspection task indication information to a low-dimensional space, so that these features can be better processed by subsequent algorithms. For example, if the target inspection task indication information contains multiple task types (such as equipment inspection, area inspection, etc.), task priorities (expressed as numerical values, such as between 0-1, 0 is the lowest priority, 1 is the highest priority) and task execution time requirements (in minutes, such as completing an area inspection within 30 minutes) and other features, these different types of features can be converted into a low-dimensional vector representation through feature embedding. For example, the features such as task type, priority and execution time requirements are converted into a vector V = [v1, v2, v3, v4, v5] with a dimension of n (such as n = 5) through a certain mapping function f, where each vi is a numerical value after mapping.

[0097] Then, the target inspection task indication information after feature embedding processing is fused with the original processing object to obtain the target inspection task fusion information. This fusion process can be carried out in a variety of ways, such as direct vector concatenation or weighted summation. For example, the original processing object is represented by a vector O = [o1, o2, o3, ..., om], and the target inspection task fusion information vector F = [o1, o2, o3, ..., om, v1, v2, v3, v4, v5] is obtained by direct concatenation.

[0098] The target inspection task fusion information is then input into the corresponding drone tracking and monitoring algorithm. The drone tracking and monitoring algorithm is an algorithm used to monitor the drone state and generate relevant state vectors. It includes long short-term memory modules (LSTM) corresponding to multiple collaborative competitive discrimination branches of the multi-agent deep deterministic policy gradient algorithm. The long short-term memory module has advantages in processing sequence data and can effectively capture the relationship between the drone state and time.

[0099] In the above technical solution, the debugging process of the drone tracking and monitoring algorithm is an important link to ensure the accuracy of the algorithm. First, the cooperative competition discrimination branch in the multi-agent deep deterministic policy gradient algorithm is iterated to obtain the iterated drone tracking and monitoring algorithm, and the algorithm variables of the multi-agent deep deterministic policy gradient algorithm are locked. This process aims to keep certain key parameters of the multi-agent deep deterministic policy gradient algorithm unchanged when debugging the drone tracking and monitoring algorithm, so as to focus more on the optimization of the drone tracking and monitoring algorithm.

[0100] Next, a second a priori debugging example set is obtained, wherein the second a priori debugging example set includes at least one historical inspection record. These historical inspection records are data sets related to inspection tasks that actually occurred in the past, contain rich information, and can be used for debugging and optimizing algorithms.

[0101] The historical multi-agent scheduling strategy, historical inspection task indication information and historical inspection quantization vector are extracted from the historical inspection records. The historical multi-agent scheduling strategy records the task scheduling method of drones in the multi-agent system in the past, such as which drones are assigned to which inspection areas, which tasks are performed, and the order in which tasks are executed; the historical inspection task indication information is similar to the current target inspection task indication information, including the task type and priority at that time; the historical inspection quantization vector is a vector that quantifies the historical inspection tasks, which may include quantitative indicators such as task completion and resource consumption.

[0102] Then, the multi-agent deep deterministic policy gradient algorithm is used to implement several rounds of perturbation configuration processing on the historical multi-agent scheduling strategy to obtain the multi-agent perturbation scheduling strategy, and cache the first perturbation feature of each perturbation configuration processing configured in the historical multi-agent scheduling strategy. Perturbation configuration processing is a purposeful modification operation on the historical multi-agent scheduling strategy, such as changing the task allocation of certain drones, adjusting the order of task execution, etc. For example, in a round of perturbation configuration processing, the inspection task of area A originally in charge of a drone is changed to the inspection task of area B. This change is a perturbation configuration, and the features related to this change (such as area change, task type change, etc.) are cached as the first perturbation feature.

[0103] The historical inspection quantization vector is input into the multi-agent deep deterministic policy gradient algorithm, so that the multi-agent deep deterministic policy gradient algorithm implements several rounds of disturbance detection processing on the multi-agent disturbance scheduling strategy based on the historical inspection quantization vector. In each round of disturbance detection processing, the historical inspection task indication information is input into the drone tracking and monitoring algorithm to obtain the drone cluster state vector sample corresponding to each collaborative competition discrimination branch. For example, for a certain collaborative competition discrimination branch, based on the input historical inspection task indication information, the drone tracking and monitoring algorithm may output a vector sample [q1, q2, q3, ..., qk] containing the status of multiple drones (such as location, power, task execution progress, etc.), where qi represents the relevant status information of the i-th drone.

[0104] According to the UAV cluster state vector sample, the output of the corresponding cooperative competition discrimination branch is optimized to obtain the second disturbance feature corresponding to the single-round disturbance detection processing. This optimization process can be performed according to a specific objective function. For example, if the goal is to minimize the task completion time, the output of the cooperative competition discrimination branch can be adjusted according to the information such as the UAV position and task execution progress in the UAV cluster state vector sample to shorten the task execution time. The feature related to this adjustment is used as the second disturbance feature.

[0105] The second network debugging error is determined based on the second disturbance feature and the first disturbance feature processed by the corresponding disturbance configuration. For example, the second network debugging error E can be calculated by the formula E=Σ(|d2i-d1i|), where d2i is the i-th element in the second disturbance feature and d1i is the i-th element in the corresponding first disturbance feature. This formula represents the calculation of the sum of the element differences between the two disturbance features, which is used to measure the error in the disturbance processing process.

[0106] According to the second network debugging error, the algorithm variables of each long short-term memory module are optimized until the UAV tracking monitoring algorithm meets the first monitoring quality requirement, thereby obtaining the UAV tracking monitoring algorithm. The first monitoring quality requirement can be a set error threshold or other evaluation criteria, such as requiring the second network debugging error to be less than a certain value (such as 0.1). When this requirement is met, it is considered that the UAV tracking monitoring algorithm has achieved acceptable accuracy and the debugging process is completed.

[0107] Therefore, by fusing the target inspection task indication information with the original processing object of the multi-agent deep deterministic policy gradient algorithm, and processing it through the UAV tracking and monitoring algorithm to generate the first UAV cluster state vector, it is possible to fully utilize the task-related features in the target inspection task indication information, combined with the basic information of the multi-agent system, to accurately reflect the state of the UAV cluster. During the debugging process of the UAV tracking and monitoring algorithm, the historical inspection records are used for disturbance configuration and detection processing, and the algorithm variables of the long short-term memory module are optimized by calculating the second network debugging error, ensuring the accuracy and reliability of the UAV tracking and monitoring algorithm. This helps to improve the monitoring ability of the entire inspection system on the state of the UAV cluster, thereby optimizing the task scheduling and path planning of the UAV, improving the inspection efficiency, reducing resource waste, and being able to better adapt to different inspection task requirements.

[0108] In some examples, the historical inspection record is a historical inspection image or a three-dimensional annotation model of the inspection space area, the historical inspection quantization vector includes one of a multi-machine nest distribution relationship vector sample and a historical inspection image descriptor, the multi-machine nest distribution relationship vector sample is obtained by residually connecting the three-dimensional annotation model of the inspection space area through a fully convolutional neural network, and the historical inspection image descriptor is obtained by residually connecting the three-dimensional annotation model of the inspection space area or the historical inspection image through a wavelet analysis network, and the fully convolutional neural network and the wavelet analysis network are jointly debugged through the following steps: obtaining a third a priori debugging example set, wherein the third a priori debugging example set includes multiple debugging example binary groups, and the debugging example binary groups include corresponding inspection image samples and three-dimensional annotation model samples; performing residual connections on each of the inspection image samples and each of the three-dimensional annotation model samples through the wavelet analysis network and the fully convolutional neural network, respectively, to obtain multiple inspection image wavelet description samples and multiple prior The invention discloses a method for testing the multi-machine-nest distribution relationship vector; the inspection image samples and the three-dimensional annotation model samples in each debugging example binary group are respectively decomposed into a plurality of corresponding inspection task path samples and a plurality of past inspection space area samples, and the plurality of inspection task path samples and the plurality of past inspection space area samples are respectively residually connected through the wavelet analysis network and the fully convolutional neural network to obtain a plurality of path planning spatiotemporal data samples and a plurality of multi-machine-nest distribution relationship training sub-vectors; the spatial attention confidence is determined according to the plurality of inspection image wavelet description samples and the plurality of prior multi-machine-nest distribution relationship vectors, and the temporal attention confidence is determined according to the plurality of path planning spatiotemporal data samples and the plurality of multi-machine-nest distribution relationship training sub-vectors; the global attention confidence is determined according to the spatial attention confidence and the temporal attention confidence, and the wavelet analysis network and the fully convolutional neural network are optimized according to the global attention confidence until the global attention confidence is greater than the preset attention confidence.

[0109] In this technical solution, the historical inspection records are in the form of historical inspection images or three-dimensional annotation models of inspection space areas. These two forms provide an important data basis for subsequent operations.

[0110] The historical inspection quantization vector includes a multi-machine nest distribution relationship vector sample and one of the historical inspection image descriptors. For the multi-machine nest distribution relationship vector sample, it is obtained by residually connecting the three-dimensional annotation model of the inspection space area through a fully convolutional neural network. When processing image data, the fully convolutional neural network (FCN) can extract features while retaining spatial information. For example, in a specific model structure, the fully convolutional neural network contains multiple convolutional layers, and the convolution kernel size of each convolutional layer is 3*3, the step size is 1, and the padding is 1. When performing residual connection on the three-dimensional annotation model of the inspection space area, taking the output feature map of a certain intermediate layer as an example, assume its size is H*W*C (H is the height, W is the width, and C is the number of channels). The residual connection operation is to add the input and output of the layer, which helps to alleviate the gradient vanishing problem and improve the feature representation capability.

[0111] The historical inspection image descriptor is obtained by residual connection of the inspection space area 3D annotation model or historical inspection image through the wavelet analysis network. The wavelet analysis network uses the characteristics of wavelet transform to perform multi-scale analysis on the image or model. For example, the wavelet function can use Haar wavelet. For the input image or model data, through wavelet decomposition and reconstruction at different scales, descriptors with different frequency characteristics are extracted. In the residual connection process, the input and output of a certain stage are also added to enhance the feature expression.

[0112] The full convolutional neural network and wavelet analysis network are optimized through joint debugging. First, the third prior debugging example set is obtained, which contains multiple debugging example tuples, each of which consists of a corresponding inspection image sample and a 3D annotation model sample.

[0113] Then, residual connections are performed on each inspection image sample and each 3D annotation model sample through the wavelet analysis network and the fully convolutional neural network, respectively, to obtain multiple inspection image wavelet description samples and multiple prior multi-machine nest distribution relationship vectors. For each inspection image sample, after residual connection processing by the wavelet analysis network, a wavelet description sample vector of the inspection image with a specific dimension is obtained, for example, its dimension is D1. For each 3D annotation model sample, after residual connection by the fully convolutional neural network, a prior multi-machine nest distribution relationship vector is obtained, for example, its dimension is D2.

[0114] Next, the inspection image samples and 3D annotation model samples in each debugging example tuple are decomposed into multiple corresponding inspection task path samples and multiple past inspection space area samples. For example, for an inspection image sample, if it represents an inspection scene of a power facility, it can be decomposed into multiple inspection task path samples according to different devices and areas in the image, such as the inspection path for a specific power tower, the inspection path for a certain section of the transmission line, etc. For the 3D annotation model samples, they are also decomposed into past inspection space area samples according to similar logic.

[0115] Then, residual connections are performed on multiple inspection task path samples and multiple past inspection spatial area samples through wavelet analysis network and full convolution neural network, respectively, to obtain multiple path planning spatiotemporal data samples and multiple multi-machine nest distribution relationship training sub-vectors. For example, for an inspection task path sample, the path planning spatiotemporal data sample obtained after residual connection of wavelet analysis network is a vector containing position, time and other information, such as [x1, y1, t1, ...], where x1 and y1 represent plane position coordinates, t1 represents time point, etc. For the multi-machine nest distribution relationship training sub-vector obtained by residual connection of past inspection spatial area samples through full convolution neural network, for example, it represents the distance, relative position and other relationships between machine nests, such as [d1, θ1, ...], where d1 represents the distance between machine nests, θ1 represents the relative azimuth, etc.

[0116] The spatial attention confidence is determined based on multiple inspection image wavelet description samples and multiple prior multi-machine nest distribution relationship vectors. The spatial attention confidence reflects the degree of correlation and reliability between features in the spatial dimension. For example, a calculation method based on cosine similarity is adopted. For the i-th inspection image wavelet description sample vector Vi and the j-th prior multi-machine nest distribution relationship vector Uj, the spatial attention confidence Si,j can be calculated by the formula Si,j = cos(Vi,Uj) = (Vi·Uj) / (||Vi||*||Uj||), where Vi·Uj represents the dot product of the vectors, and ||Vi|| and ||Uj|| represent the modulus of the vectors Vi and Uj respectively. This value ranges from -1 to 1, and the closer it is to 1, the higher the spatial attention confidence.

[0117] The temporal attention confidence is determined based on multiple path planning spatiotemporal data samples and multiple multi-machine nest distribution relationship training sub-vectors. The temporal attention confidence reflects the closeness and reliability of the correlation between data in the time series. For example, the temporal attention confidence is determined by calculating the dynamic time warping (DTW) distance between the path planning spatiotemporal data samples and the multi-machine nest distribution relationship training sub-vectors. Suppose the path planning spatiotemporal data samples are the sequence A = [a1, a2, ..., an], the multi-machine nest distribution relationship training sub-vectors are the sequence B = [b1, b2, ..., bm], and the distance calculated by the DTW algorithm is D(A, B), then the temporal attention confidence Ti,j = 1 / (1+D(A, B)), the value range is between 0 and 1, the closer to 1, the higher the temporal attention confidence.

[0118] The global attention confidence is determined based on the spatial attention confidence and the temporal attention confidence. For example, the global attention confidence G can be calculated by weighted summation, that is, G = α*S + β*T, where S is a comprehensive measure of the spatial attention confidence (such as the average spatial attention confidence), T is a comprehensive measure of the temporal attention confidence (such as the average temporal attention confidence), and α and β are weight coefficients used to adjust the contribution ratio of the spatial and temporal attention confidences to the global attention confidence.

[0119] The wavelet analysis network and the full convolutional neural network are optimized according to the global attention confidence until the global attention confidence is greater than the preset attention confidence. This preset attention confidence is a preset threshold, such as 0.8. During the optimization process, if the global attention confidence is less than the preset attention confidence, the parameters of the wavelet analysis network and the full convolutional neural network, such as the weight of the convolution kernel, the parameters of the wavelet function, etc., are adjusted, and then the above calculations and processing are performed again until the global attention confidence is greater than the preset attention confidence, thereby ensuring that the performance of the wavelet analysis network and the full convolutional neural network meets the requirements.

[0120] In this way, by obtaining multi-machine nest distribution relationship vector samples and historical inspection image descriptors in the historical inspection quantization vector in a specific way, effective information in the inspection data can be mined from different angles. The linkage debugging process of the full convolutional neural network and the wavelet analysis network, using the calculation and optimization of spatial attention confidence, temporal attention confidence and global attention confidence, can effectively improve the network's feature extraction capabilities for inspection images and three-dimensional annotation models. This helps to analyze inspection data more accurately, optimize subsequent inspection task planning, and improve the adaptability between inspection tasks and machine nest distribution relationships and inspection paths, thereby improving the efficiency and accuracy of the entire inspection system.

[0121] In some scalable technical ideas, after obtaining the target inspection task indication information of the target inspection space area, the method also includes: obtaining the multi-level inspection task coverage indication of the target inspection space area. Then, the autonomous collaborative inspection strategy for the target inspection space area is generated based on the second collaborative inspection scheduling viewpoint, so that the autonomous collaborative inspection strategy contains the inspection task path corresponding to the target inspection task indication information, including: generating a second drone cluster state vector corresponding to the collaborative competition discrimination branch based on the multi-level inspection task coverage indication; optimizing the second collaborative inspection scheduling viewpoint based on the second drone cluster state vector to obtain a third drone cluster state vector; generating the autonomous collaborative inspection strategy for the target inspection space area based on the third drone cluster state vector, so that the autonomous collaborative inspection strategy contains the inspection task path corresponding to the target inspection task indication information according to the multi-level coverage annotation corresponding to the multi-level inspection task coverage indication.

[0122] In detail, the method of obtaining the multi-level inspection task coverage indication of the target inspection space area includes: marking the target inspection task label from the target inspection space area; obtaining the task cycle of the target inspection task label; obtaining the task condition of the target inspection task label; if the task cycle and the task condition meet the target monitoring quality requirements, obtaining the multi-level inspection task coverage indication of the target inspection space area.

[0123] In addition, after marking the target inspection task label from the target inspection space area, the method also includes: if the target inspection task label is an efficiency priority task, identifying the efficiency planning route node set in the inspection space sub-area corresponding to the efficiency priority task in the target inspection space area; if the efficiency planning route node set meets the efficiency compliance requirements, obtaining a multi-level inspection task coverage indication for the target inspection space area.

[0124] It is understandable that after obtaining the target inspection task indication information of the target inspection space area, it is also involved to obtain the multi-level inspection task coverage indication of the target inspection space area. This operation is of great significance for generating an autonomous collaborative inspection strategy for the target inspection space area based on the second collaborative inspection scheduling viewpoint.

[0125] First, the specific process of obtaining the multi-level inspection task coverage indication of the target inspection space area is explained. The target inspection task label is marked from the target inspection space area. This marking process needs to be based on specific rules or algorithms. For example, for a complex inspection space area containing multiple equipment and areas (such as a large industrial plant), it can be marked according to factors such as equipment type and area function. For example, if there are different types of areas such as production equipment, storage facilities, and transportation channels in the plant, then the target inspection task label corresponding to the production equipment may be "equipment operation status monitoring", and the target inspection task label for the storage facility may be "cargo storage safety monitoring", etc.

[0126] Next, the task cycle of the target inspection task label is obtained. The task cycle refers to the time interval for executing the inspection task, measured in time units. For example, for the target inspection task label "equipment operation status monitoring", its task cycle may be an inspection every 2 hours; for "cargo storage safety monitoring", it may be once every 4 hours. The determination of the task cycle may be based on various factors such as the importance of the equipment, the stability of the equipment operation, the nature of the cargo, etc.

[0127] At the same time, the task conditions of the target inspection task label are obtained. The task conditions include various conditions that need to be met to execute the inspection task, such as environmental conditions (temperature, humidity, etc.), equipment operating status (on, off, etc.), etc. For example, for the "equipment operating status monitoring" task of some high-precision equipment, the task condition may be that the equipment is running and the ambient temperature is between 20-25 degrees Celsius.

[0128] If the task cycle and task conditions meet the target monitoring quality requirements, a multi-level inspection task coverage indication for the target inspection space area is obtained. The target monitoring quality requirements are a set of pre-set standards used to measure whether the task cycle and task conditions can meet the needs of effective inspection. For example, the target monitoring quality requirements may stipulate that the task cycle cannot exceed the longest non-inspection time interval allowed for the safe operation of the equipment, and the fluctuation range of the environmental conditions in the task conditions cannot exceed the environmental parameter range for the normal operation of the equipment.

[0129] In addition, after marking the target inspection task label from the target inspection space area, if the target inspection task label is an efficiency priority task, further operations are required. Identify the efficiency planning route node set in the inspection space sub-area corresponding to the efficiency priority task in the target inspection space area. For example, in a large-scale power inspection scenario, if "inspection of key components of power towers" is marked as an efficiency priority task, then the inspection space sub-area corresponding to this task is the area where each power tower is located, and the efficiency planning route node set may be the location points of key components on the power tower (such as insulators, lightning arresters, etc.). The determination of these node sets needs to take into account factors such as the efficiency of drone flight and the comprehensiveness of inspections.

[0130] If the efficiency planning route node set meets the efficiency standard requirements, the multi-level inspection task coverage indication of the target inspection space area is obtained. The efficiency standard requirements are also pre-set standards, such as requiring the efficiency planning route node set to cover more than 90% of key components and the total flight distance or flight time of the drone when flying according to the route node set is the shortest.

[0131] In terms of generating an autonomous collaborative inspection strategy for the target inspection space area based on the second collaborative inspection scheduling viewpoint, so that the autonomous collaborative inspection strategy includes the inspection task path corresponding to the target inspection task indication information, firstly, the second drone cluster state vector corresponding to the collaborative competition discrimination branch is generated based on the multi-level inspection task coverage indication. This process needs to comprehensively consider various information in the multi-level inspection task coverage indication, such as the coverage of inspection tasks at different levels, the priority of different tasks, etc. For example, the multi-level inspection task coverage indication includes a first-level inspection task (such as inspection of key equipment) and a second-level inspection task (such as inspection of ordinary equipment). The coverage of the first-level inspection task is smaller but the priority is higher, and the coverage of the second-level inspection task is larger but the priority is lower. When generating the second drone cluster state vector, it is necessary to determine the state information such as the number of drones allocated and the initial position based on this information.

[0132] Then, the second collaborative inspection scheduling viewpoint is optimized based on the second drone cluster state vector to obtain the third drone cluster state vector. This optimization process may involve adjustments to the drone's task allocation, flight path planning, and other aspects. For example, if the second drone cluster state vector shows that there are too many drones in a certain area, while the key tasks in another area are not covered enough, then when optimizing the second collaborative inspection scheduling viewpoint, the drone's task allocation will be adjusted so that more drones go to the area with insufficient coverage of the key tasks, thereby obtaining the third drone cluster state vector.

[0133] Finally, the autonomous collaborative inspection strategy for the target inspection space area is generated based on the state vector of the third drone cluster, so that the multi-level coverage annotation corresponding to the multi-level inspection task coverage indication in the autonomous collaborative inspection strategy contains the inspection task path corresponding to the target inspection task indication information. When generating the autonomous collaborative inspection strategy, it is necessary to ensure that the inspection task path can meet the requirements in the multi-level inspection task coverage indication. For example, if the multi-level coverage annotation requires the completion of the first-level inspection task before the second-level inspection task, then in the autonomous collaborative inspection strategy, the flight path of the drone will first point to the area corresponding to the first-level inspection task, and after completing these tasks, it will go to the area corresponding to the second-level inspection task, and these task paths must match the task requirements (such as task type, task cycle, etc.) in the target inspection task indication information.

[0134] By applying the above technical ideas and obtaining multi-level inspection task coverage indications, it is possible to plan inspection tasks more carefully and meet inspection needs at different levels. Determining multi-level inspection task coverage indications based on whether the task cycle and task conditions meet the target monitoring quality requirements and whether the efficiency planning route node set meets the efficiency requirements will help ensure the effectiveness and efficiency of the inspection tasks. In the process of generating autonomous collaborative inspection strategies, a series of vector generation and optimization operations are performed based on multi-level inspection task coverage indications, which can reasonably allocate drone resources, optimize flight paths, improve the accuracy and comprehensiveness of inspections, and ensure the priority execution of key tasks, thereby improving the efficiency of the entire inspection system.

[0135] In some other scalable technical ideas, the second collaborative inspection scheduling concept is optimized according to the second drone cluster state vector to obtain a third drone cluster state vector, including: obtaining a first state change trend characteristic value corresponding to the first drone cluster state vector, and a second state change trend characteristic value corresponding to the second drone cluster state vector, wherein the sum of the first state change trend characteristic value and the second state change trend characteristic value is not greater than a set characteristic value; strengthening the first drone cluster state vector and the second drone cluster state vector according to the first state change trend characteristic value and the second state change trend characteristic value respectively to obtain a strengthened first drone cluster state vector and a strengthened second drone cluster state vector, and based on the strengthened first drone cluster state vector and the strengthened second drone cluster state vector, and the first collaborative inspection scheduling concept of the corresponding collaborative competition judgment branch, the third drone cluster state vector is obtained.

[0136] In the framework of the autonomous inspection system for drones with multi-nest take-off and landing, an exemplary introduction to the technical solution of optimizing the second collaborative inspection scheduling concept based on the second drone cluster state vector to obtain the third drone cluster state vector is described as follows.

[0137] First, about obtaining the first state change trend characteristic value and the second state change trend characteristic value. The first drone cluster state vector contains a lot of key information. For example, if there are 10 drones participating in a specific inspection task, the first drone cluster state vector may contain the real-time position coordinates of each drone (expressed in longitude and latitude, with an accuracy of 6 decimal places, such as the first drone is located at 30.123456° north latitude and 120.123456° east longitude), the remaining power of each drone (in milliampere hours mAh, for example, the remaining power of the first drone is 5000mAh), the proportion of each drone's completed inspection tasks to the total tasks (value range 0-1, for example, the first drone has completed 0.3 of the inspection tasks) and other information.

[0138] When calculating the first state change trend characteristic value, for the position coordinates, if the change within a specific time period (such as 10 minutes) is considered. Assume that the initial position coordinates of the first drone are ((x_1, y_1)), and the position coordinates after 10 minutes are ((x_2, y_2)), the position change rate in the (x) axis direction (r_x = frac{x_2-x_1}{10}) (unit: degree / minute), for the entire drone cluster, the position change rates of all drones in the (x) axis direction are weighted and summed according to a certain weight (such as assigning weights according to the importance of the drone's mission, and important drones have higher weights), and the position trend value (Tr_x) of the entire cluster in the (x) axis direction is obtained; similarly, the position trend value (Tr_y) in the (y) axis direction is obtained.

[0139] In terms of power consumption, assume that the total power consumption of all drones at the initial moment is (E_1) (unit: mAh), the total power consumption after 10 minutes is (E_2) (unit: mAh), and the power consumption trend value is (Tr_E = frac{E_2-E_1}{10}) (unit: mAh / minute).

[0140] As for the task completion ratio, assume that the sum of the task completion ratios of all drones at the initial moment is (P_1), and 10 minutes later is (P_2), and the task ratio trend value (Tr_P = frac{P_2-P_1}{10}) (unit: none, value range -1 to 1, positive value indicates an increase in the task completion ratio, and negative value indicates a decrease).

[0141] The first state change trend characteristic value:

[0142] (Tr_1=w_xTr_x+w_yTr_y+w_ETr_E+w_PTr_P), where (w_x), (w_y), (w_E), (w_P) are weight coefficients, and (w_x+w_y+w_E+w_P=1), these weight coefficients are set according to different inspection task requirements and system priorities. For example, for long-distance inspection tasks with high power requirements, the value of (w_E) may be higher; for emergency inspection tasks with high task completion speed requirements, the value of (w_P) may be higher.

[0143] Similarly, for the second drone cluster state vector, the second state change trend characteristic value (Tr_2) is calculated according to the same logic, and (Tr_1+Tr_2leqS) ((S) is the set characteristic value, which is determined based on multiple factors such as the overall performance of the system and resource constraints. For example, it is set as a comprehensive value based on the maximum state change rate that the system can withstand, such as (S=0.5)).

[0144] Next, the first drone cluster state vector and the second drone cluster state vector are enhanced according to the first state change trend characteristic value and the second state change trend characteristic value, respectively.

[0145] For the enhanced processing of the state vector of the first drone cluster, if (Tr_1>0), it indicates that the overall situation is in a positive development trend. Taking the power as an example, if (Tr_E<0), it means that the power is decreasing, and there are charging facilities in the system, and the charging power is (P_c) (unit: mAh / minute), and the charging strategy is adjusted according to the value of (Tr_E). For example, according to (△E=-ktimesTr_E) ((k) is the adjustment coefficient, which is determined according to the performance of the charging facilities and the battery characteristics, such as (k=2)), the charging amount is increased, that is, the adjusted charging amount is (P_c'=P_c+△E).

[0146] In terms of position, if (Tr_x>0) and in the target inspection direction (for example, the target inspection area is in the positive direction of the (x) axis), for each drone, for example, its basic flight speed is (v_0) (unit: meters / second), the flight speed can be increased according to (v_x=v_0+mtimesTr_x) ((m) is the speed adjustment coefficient, determined according to the drone performance and aerodynamic characteristics, for example (m=10)) to reach the target area faster.

[0147] For the task ratio, if (Tr_P<1), it means that the task completion ratio is growing slowly. For example, there are 10 task types in the system, and the task allocation matrix is ​​(M=(m_{ij})), where (i=1, 2, cdots, 10) represents the task type, (j=1, 2, cdots, 10) represents the drone number, and (m_{ij}) represents the priority of the (j)th drone to perform the (i)th task (value range 0-1). If it is found that the completion ratio of a certain type of task (such as the third type of task) is growing slowly, and the drone that performs the task (such as the fifth drone) has sufficient power and is in a suitable position, the value of (m_{35}) can be increased, such as (m_{35}=m_{35}+△m) ((△m) is determined according to the specific situation, for example (△m=0.1)), so that the fifth drone is more inclined to perform the third type of task, thereby accelerating the growth of the task completion ratio.

[0148] The second UAV cluster state vector is enhanced in the same manner to obtain an enhanced first UAV cluster state vector and an enhanced second UAV cluster state vector.

[0149] Finally, the third UAV cluster state vector is obtained based on the enhanced first UAV cluster state vector and the enhanced second UAV cluster state vector, as well as the first collaborative inspection scheduling viewpoint of the corresponding collaborative competition discrimination branch. The first collaborative inspection scheduling viewpoint contains a lot of information, such as task allocation strategy, collaborative and competitive relationships between UAVs, etc.

[0150] For example, in a specific inspection scenario, there are 5 drones participating in the inspection, and the task allocation strategy specifies the inspection area and task type for each drone. Under the enhanced state vector, if it is found that a certain area (such as area A) was originally inspected by 2 drones, but after enhancement, it is found that the resources (such as power, flight speed, etc.) of these 2 drones are still surplus, and the inspection task of another area (such as area B) is heavier, under the first collaborative inspection scheduling point of view, the task allocation can be adjusted to transfer one of the drones from area A to area B. This adjustment may involve operations such as changing the corresponding element values ​​in the task allocation matrix and replanning the flight path of the drone. The third drone cluster state vector is constructed based on this adjusted information. This vector will be more in line with the actual needs of the current inspection task, including the reasonable distribution of drones and the efficient execution of tasks.

[0151] It can be seen that by accurately calculating the characteristic values ​​of the state change trend, the dynamic changes of the state of the drone cluster can be carefully grasped. According to these characteristic values, the enhanced processing can be carried out to timely adjust the strategy for the problems of the drone cluster in terms of location, power, task completion ratio, etc., so that the state of the drone cluster develops in a direction that is more conducive to the completion of the inspection task. Based on the enhanced state vector optimization collaborative inspection scheduling point of view to obtain the third drone cluster state vector, it can achieve a more reasonable allocation of drone resources and improve the inspection efficiency. This is combined with the advantages of the drone autonomous inspection system with multi-machine nest take-off and landing in different places, such as rapid response to tasks without time and region restrictions, automatic inspection without manual intervention, and flexibility and coverage improvement brought by the multi-machine nest layout, etc., which further enhances the performance of the entire inspection system and ensures that the inspection task can be completed efficiently, accurately and comprehensively.

[0152] From the above content, we can see that in the autonomous inspection system of drones with multiple take-off and landing locations, there are a series of technical solutions to achieve its significant application effects.

[0153] The autonomous inspection system of unmanned aerial vehicles can take off from the nearest nest at any time to perform tasks according to preset instructions, and is not restricted by time and region, thereby improving response speed and work efficiency. There is corresponding technical support behind this feature. In the entire system architecture, there must be an efficient instruction receiving and processing module, which can obtain preset instructions in real time and parse them quickly. For example, the system receives an instruction to inspect a specific area, and the instruction contains the coordinate information or specific identification of the target area. The system parses the instruction in a very short time (for example, within 1-2 seconds) to determine the target area. At the same time, the system has a fast nest selection algorithm, which compares the location information of each nest (accurately expressed in longitude and latitude coordinates, such as the coordinates of nest 1 are 30° north latitude and 120° east longitude; the coordinates of nest 2 are 31° north latitude and 121° east longitude, etc.) with the location of the target area to calculate the distance (which can be calculated according to the distance formula between two points), and selects the nest closest to the target area. This fast instruction processing and nest selection mechanism greatly shortens the task preparation time and improves the response speed. And since no additional manpower is required to arrange matters such as the take-off location, the intermediate links are reduced and operational efficiency is improved.

[0154] The system does not require manual intervention throughout the process. Multiple drones can take off with one button to automatically conduct full-process aerial inspections, and the operating efficiency is more than 4 times higher than that of manual inspections. The system has a complex and orderly automated process. On the drone side, each drone is equipped with a high-performance flight control system. This system integrates advanced flight control algorithms and can accurately control the takeoff, flight attitude, route tracking and other operations of the drone. For example, when the drone takes off, the flight control system accurately controls the motor speed and rudder operation according to the preset takeoff parameters (such as a takeoff speed of 5 meters per second, a takeoff angle of 30° with the horizontal direction, etc.) to ensure a smooth takeoff. During the flight of the route, the high-precision positioning system (such as GPS positioning accuracy of up to 1-3 meters) is used to obtain its own position information, and it is compared with the preset inspection route (the route consists of a series of coordinate points, such as [(x1, y1, z1), (x2, y2, z2), ...]), and the error is controlled within a very small range (for example, the error range is 1-2 meters) to ensure accurate flight along the route.

[0155] The collaborative operation of multiple drones has also been carefully designed. At the system level, there is a central coordination module, which is responsible for allocating and coordinating the tasks of multiple drones. For example, when there are 5 drones participating in the inspection task, the central coordination module divides the entire inspection area into 5 sub-areas (the area of ​​each sub-area is calculated according to the area division algorithm) according to the size of the target inspection area (for example, the target inspection area is a rectangular area with a length of 10 kilometers and a width of 5 kilometers) and the task type (such as visual inspection and temperature detection of facilities in the area), and then allocates a sub-area to each drone. During the inspection process, drones exchange information through wireless communication modules (the communication frequency is a specific frequency band, such as the 2.4GHz frequency band, and the communication rate reaches more than 1Mbps per second). For example, if a drone finds an abnormal situation (such as a certain device temperature is too high, the temperature value reaches 80℃, and the normal temperature range is 30-50℃), it will quickly send this information to other drones and the central coordination module through the wireless communication module. The central coordination module may adjust the inspection focus or task priority of other drones based on this information. This automated collaborative process is much more efficient than manual inspection. Taking a specific inspection task as an example, if a large factory (factory area is 1 million square meters) is inspected manually, it will take 20 workers a whole day (8 hours) to complete the initial inspection. However, using this drone autonomous inspection system, 5 drones may be able to complete the inspection task of the same scope within 2 hours, which is more than 4 times more efficient.

[0156] In the core inspection area, multiple nests are arranged to form an efficient and coordinated triangular layout to improve the flexibility and coverage of the inspection. The triangular layout has unique geometric advantages. For example, in a large inspection area (such as a nature reserve with an area of ​​50 square kilometers), three nests are arranged to form a triangular layout. The distances between the three vertex nests of the triangle have been carefully designed. For example, the distance between nest A and nest B is 10 kilometers, the distance between nest B and nest C is 12 kilometers, and the distance between nest A and nest C is 15 kilometers. This layout allows the drone to cover a larger area when taking off from any nest. When the drone takes off from nest A to inspect the area close to nest B, due to the existence of nest C, a cross-coverage effect can be formed to avoid inspection blind spots. Moreover, this layout is highly flexible in responding to different inspection task requirements. If at a certain moment it is necessary to focus on inspecting an area near a certain edge in the triangle layout (such as a newly discovered object that needs special attention near the edge of Machine Nest A-Machine Nest B), the drones located at Machine Nest A and Machine Nest B can be quickly dispatched to increase the inspection frequency and depth without affecting the basic inspection tasks in other areas.

[0157] It can be seen that the response speed is improved and the task preparation time is reduced through efficient command processing and nest selection mechanism. The automated flight control and multi-machine collaborative operation of the UAV improve the accuracy and efficiency of the inspection, greatly shorten the inspection time, and can quickly respond to abnormal situations during the inspection. The triangular layout of multiple nests increases the flexibility and coverage of the inspection, ensures a comprehensive inspection of the target area, and avoids inspection loopholes. These features work together to make the entire UAV autonomous inspection system show significant advantages in various inspection tasks.

[0158] Further, Figure 2 The structure diagram of a drone autonomous inspection system 200 provided in an embodiment of the present application is shown in FIG. Figure 2 The drone autonomous inspection system 200 shown includes a processor 210, which can call and run a computer program from a memory to implement the method in the embodiment of the present application. Figure 2 As shown, the drone autonomous inspection system 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present application. The memory 230 may be a separate device independent of the processor 210, or may be integrated in the processor 210. Optionally, as Figure 2As shown, the drone autonomous inspection system 200 may also include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices. Optionally, the drone autonomous inspection system 200 may implement the corresponding processes corresponding to the storage engine or the components in the storage engine (such as a processing module) or the device deployed with the storage engine in each method of the embodiment of the present application. For the sake of brevity, it will not be repeated here. It should be understood that the processor of the embodiment of the present application may be an integrated circuit chip with signal processing capabilities. It can be understood that the memory in the embodiment of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. It should be noted that the memory of the system and method described herein is intended to include, but is not limited to, a memory of a suitable type.

[0159] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0160] The embodiments of the embodiments of the present application are described above in conjunction with the accompanying drawings, but the embodiments of the present application are not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the embodiments of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the embodiments of the present application and the scope of protection of the embodiments of the present application, all of which are within the protection of the embodiments of the present application.

Claims

1. A multi-host remote take-off and landing UAV autonomous inspection method, characterized by: The method is applied to a UAV autonomous inspection system, and the method includes: Identify the target inspection space area, obtain the multi-drone distribution relationship vector corresponding to the target inspection space area, and obtain target inspection task indication information of the target inspection space area; the target inspection task indication information includes the collaborative inspection task requirement vector of multiple drones; Inputting the multi-hive distribution relationship vector into a multi-agent deep deterministic policy gradient algorithm, wherein the multi-agent deep deterministic policy gradient algorithm includes a collaborative competitive discriminant branch for performing discriminant analysis on the spatiotemporal data of the UAV path planning based on the multi-hive distribution relationship vector; Based on the target inspection task indication information, a first UAV cluster state vector corresponding to the collaborative competition judgment branch is generated, and based on the first UAV cluster state vector, the first collaborative inspection scheduling viewpoint of the collaborative competition judgment branch is optimized to obtain a second collaborative inspection scheduling viewpoint, and based on the second collaborative inspection scheduling viewpoint, an autonomous collaborative inspection strategy for the target inspection space area is generated so that the autonomous collaborative inspection strategy includes an inspection task path corresponding to the target inspection task indication information.

2. The autonomous inspection method for UAVs with multiple take-off and landing locations according to claim 1, characterized in that: The obtaining of target inspection task indication information of the target inspection space area includes: Locating the inspection task tag from the target inspection space area, and obtaining the target inspection space sub-area containing the inspection task tag; Inputting the target inspection spatial sub-area and the associated spatial sub-area of ​​the target inspection spatial sub-area in the target inspection spatial area into a task priority parsing network to obtain the parsed task priority of the inspection task label; Obtaining the take-off and landing time efficiency weight of the inspection task tag corresponding to the inspection task tag; determining the target inspection task tag from the located inspection task tags based on the task priority and the take-off and landing time efficiency weight of the inspection task tag; Obtain the target inspection task indication information of the target inspection task tag.

3. The autonomous inspection method for UAVs with multiple take-off and landing locations according to claim 1, characterized in that: The obtaining of target inspection task indication information of the target inspection space area includes: Dividing the target inspection space area to obtain inspection space sub-areas; Inputting the divided inspection space sub-region and the set reference feature into a task feature mining network, so that the task feature mining network generates an inspection task one-hot vector according to the inspection space sub-region, wherein the set reference feature is used to characterize the generation of the inspection task one-hot vector by the task feature mining network according to the inspection space sub-region; The inspection task one-hot vector is residually connected through a deep residual connection network to obtain the target inspection task indication information.

4. The autonomous inspection method for UAVs with multiple take-off and landing locations according to claim 3 is characterized in that: The deep residual connection network is debugged by the following steps: Obtain a first a priori debugging example set, wherein the first a priori debugging example set includes a plurality of past inspection area task tuples, each of the past inspection area task tuples includes a mutually corresponding past inspection spatial area and a past inspection task one-hot vector; Performing residual connections on each of the past inspection spatial regions through an initial convolutional neural network, and performing residual connections on each of the past inspection task one-hot vectors through the deep residual connection network, to obtain a plurality of past inspection spatial region vectors and past inspection task one-hot vectors; Identifying a first correlation coefficient between each of the past inspection spatial region vectors and each of the past inspection task one-hot vectors, and determining a second correlation coefficient between each of the past inspection spatial region vectors and each of the past inspection task one-hot vectors based on the adaptation characteristics between the past inspection spatial region and the past inspection task one-hot vector; Determine a first network debugging error based on the first correlation coefficient and the second correlation coefficient, and optimize the deep residual connection network and the initial convolutional neural network based on the first network debugging error until the first network debugging error is less than a first debugging error threshold, and terminate the debugging.

5. The autonomous inspection method for UAVs with multiple take-off and landing locations according to claim 3 is characterized in that: After performing residual connection on the inspection task one-hot vector through the deep residual connection network to obtain the target inspection task indication information, the method further includes: Obtain a first policy construction reference feature and a second policy construction reference feature, and map the first policy construction reference feature and the second policy construction reference feature into corresponding policy task matching vectors; wherein, the first policy construction reference feature is used to characterize the first template inspection task label corresponding to the autonomous collaborative inspection strategy generated by the multi-agent deep deterministic policy gradient algorithm, and the second policy construction reference feature is used to characterize the second template inspection task label that needs to be excluded by the autonomous collaborative inspection strategy generated by the multi-agent deep deterministic policy gradient algorithm; The policy task matching vector is integrated into the target inspection task indication information to adjust the target inspection task indication information.

6. The autonomous inspection method for UAVs with multiple take-off and landing locations according to claim 1, characterized in that: The discriminant analysis of the spatiotemporal data of the UAV path planning based on the multi-machine nest distribution relationship vector includes: In the cooperative competition judgment branch, determining the path planning transient spatiotemporal data according to the UAV path planning spatiotemporal data; Generate a spatiotemporal feature relationship network based on the transient spatiotemporal data of the path planning, and generate a task identification feature relationship network and an inspection point feature relationship network based on the multi-machine nest distribution relationship vector; A feature collision operation is performed based on the spatiotemporal feature relationship network, the task identification feature relationship network, and the inspection point feature relationship network, and the first collaborative inspection scheduling viewpoint corresponding to the collaborative competition judgment branch is determined based on the result of the feature collision operation.

7. The autonomous inspection method for UAVs with multiple take-off and landing locations according to claim 1, characterized in that: The discriminant analysis of the spatiotemporal data of the UAV path planning based on the multi-machine nest distribution relationship vector also includes: Obtain a first statistical value, and generate a number of initial inspection path spatiotemporal features corresponding to the first statistical value, and generate a number of UAV path planning spatiotemporal data corresponding to the first statistical value based on the initial inspection path spatiotemporal features; Inputting the spatiotemporal data of the UAV path planning corresponding to the first statistical value into a multi-agent deep deterministic policy gradient algorithm, and performing discriminant analysis on the spatiotemporal data of the UAV path planning corresponding to the first statistical value according to the multi-machine nest distribution relationship vector, so as to obtain a plurality of pending autonomous collaborative inspection strategies, wherein the number of the pending autonomous collaborative inspection strategies is the same as the first statistical value; An autonomous collaborative inspection strategy is determined from the plurality of autonomous collaborative inspection strategies to be processed.

8. The autonomous inspection method for UAVs with multiple take-off and landing locations according to claim 1, characterized in that: The step of generating a first UAV cluster state vector corresponding to the collaborative competition discrimination branch according to the target inspection task indication information includes: Obtaining the original processing object of the first collaborative competitive discrimination branch of the multi-agent deep deterministic policy gradient algorithm; Performing feature embedding processing on the target inspection task indication information, and fusing the target inspection task indication information after feature embedding processing with the original processing object to obtain target inspection task fusion information; The target inspection task fusion information is input into the corresponding drone tracking and monitoring algorithm, and feature embedding processing is performed on the monitoring results of the drone tracking and monitoring algorithm to obtain the first drone cluster state vector.

9. The autonomous inspection method for UAVs with multiple take-off and landing locations according to claim 8, characterized in that: The debugging process of the UAV tracking and monitoring algorithm includes: Iterating the cooperative competitive discrimination branch in the multi-agent deep deterministic policy gradient algorithm to obtain the iterated drone tracking and monitoring algorithm, and locking the algorithm variables of the multi-agent deep deterministic policy gradient algorithm, wherein the drone tracking and monitoring algorithm includes long short-term memory modules corresponding to multiple cooperative competitive discrimination branches of the multi-agent deep deterministic policy gradient algorithm; Acquire a second a priori debugging example set, where the second a priori debugging example set includes at least one historical inspection record; Extracting historical multi-agent scheduling strategies, historical inspection task indication information, and historical inspection quantization vectors from the historical inspection records; Performing several rounds of perturbation configuration processing on the historical multi-agent scheduling strategy using the multi-agent deep deterministic policy gradient algorithm to obtain a multi-agent perturbation scheduling strategy, and caching the first perturbation feature of each perturbation configuration process configured on the historical multi-agent scheduling strategy; The historical inspection quantization vector is input into the multi-agent deep deterministic policy gradient algorithm, so that the multi-agent deep deterministic policy gradient algorithm implements several rounds of disturbance detection processing on the multi-agent disturbance scheduling strategy according to the historical inspection quantization vector; in each round of disturbance detection processing, the historical inspection task indication information is input into the drone tracking and monitoring algorithm to obtain a drone cluster state vector sample corresponding to each collaborative competition discrimination branch, and the output of the corresponding collaborative competition discrimination branch is optimized according to the drone cluster state vector sample to obtain a second disturbance feature corresponding to a single round of disturbance detection processing, and several rounds of disturbance detection processing correspond to several rounds of disturbance configuration processing; determining a second network debugging error based on the second disturbance feature and the first disturbance feature processed by the corresponding disturbance configuration, and optimizing the algorithm variables of each of the long short-term memory modules based on the second network debugging error until the drone tracking and monitoring algorithm meets the first monitoring quality requirement, thereby obtaining the drone tracking and monitoring algorithm; The historical inspection record is a historical inspection image or a three-dimensional annotation model of an inspection space area. The historical inspection quantization vector includes one of a multi-machine nest distribution relationship vector sample and a historical inspection image descriptor. The multi-machine nest distribution relationship vector sample is obtained by performing residual connection on the three-dimensional annotation model of the inspection space area through a fully convolutional neural network. The historical inspection image descriptor is obtained by performing residual connection on the three-dimensional annotation model of the inspection space area or the historical inspection image through a wavelet analysis network. The fully convolutional neural network and the wavelet analysis network are debugged in conjunction with each other through the following steps: Obtaining a third a priori debugging example set, wherein the third a priori debugging example set includes a plurality of debugging example 2-tuples, and the debugging example 2-tuples include corresponding inspection image samples and 3D annotation model samples; Performing residual connection on each inspection image sample and each three-dimensional annotation model sample through the wavelet analysis network and the fully convolutional neural network, respectively, to obtain multiple inspection image wavelet description samples and multiple prior multi-machine nest distribution relationship vectors; Decomposing the inspection image sample and the three-dimensional annotation model sample in each debugging example binary into a plurality of corresponding inspection task path samples and a plurality of past inspection space area samples, and performing residual connection on the plurality of inspection task path samples and the plurality of past inspection space area samples respectively through the wavelet analysis network and the fully convolutional neural network, thereby obtaining a plurality of path planning spatiotemporal data samples and a plurality of multi-machine nest distribution relationship training sub-vectors; Determining spatial attention confidence based on the multiple inspection image wavelet description samples and the multiple prior multi-machine nest distribution relationship vectors, and determining temporal attention confidence based on the multiple path planning spatiotemporal data samples and the multiple multi-machine nest distribution relationship training subvectors; Determine a global attention confidence based on the spatial attention confidence and the temporal attention confidence, and optimize the wavelet analysis network and the fully convolutional neural network based on the global attention confidence until the global attention confidence is greater than a preset attention confidence.

10. An autonomous inspection system for drones, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images